Skip to main content

rete_core/
file.rs

1//! `.rete` file assembly and reading (SPEC.md §4, §9).
2//!
3//! v0 layout:
4//!
5//! ```text
6//! [0..1024)  header ([`crate::header::HEADER_LEN`])
7//! [dict]     dictionary container: 4 front-coded sections
8//! [index]    permutation container: 6 triple blocks (SPO/POS/OSP/SOP/PSO/OPS)
9//! [pyramid]  summary meta (and, in future, tile directories)
10//! [footer]   trailing magic
11//! ```
12//!
13//! The header points at the dictionary container (`dictionary_offset/len`) and
14//! the permutation container (`root_dir_offset/len`); routed readers can fetch a
15//! single permutation payload from that container.
16
17use crate::dictionary::Dictionary;
18use crate::header::{
19    Header, Section, SectionKind, FLAG_HAS_QUADS, FLAG_HAS_QUOTED_TRIPLES, FLAG_TILE_SYNOPSIS,
20    HEADER_LEN, MAGIC,
21};
22use crate::index::{GraphIndex, IndexPermutation, Pattern, PermSet, NUM_PERMS};
23use crate::meta::{ClassNode, CommunityDescriptor, LevelLinks, LevelRollup, PyramidMeta};
24use crate::pyramid::{build_dendrogram, project_graph, PyramidAlgo};
25use crate::reader::RangeReader;
26use crate::tiling::{choose_round_for_budget, summarize, SuperEdge};
27use crate::triples::Triple;
28use crate::varint::{read_uvarint, write_uvarint};
29
30#[cfg(test)]
31mod hardening_tests;
32
33/// Default per-tile byte budget `T` (SPEC.md §7.1).
34pub const DEFAULT_TILE_BUDGET: usize = 64 * 1024;
35
36/// Build the encoded pyramid-meta section for a graph: cluster, pick a round
37/// sized to `budget`, then emit the **summary** (quotient) graph. Returns
38/// `(encoded_meta, pyramid_levels)`.
39///
40/// Per-community tiles are *not* stored: they would duplicate every triple, and
41/// the exact ranged single-pattern path now routes into one permutation section
42/// without that fourth copy. Physical community-tile directories are the next
43/// storage step (SPEC §7.2).
44pub fn build_pyramid_meta(
45    dict: &Dictionary,
46    triples: &[(u32, u32, u32)],
47    budget: usize,
48) -> (Vec<u8>, u16) {
49    build_pyramid_meta_with(dict, triples, budget, None)
50}
51
52/// Like [`build_pyramid_meta`], but `type_override` forces the schema-pyramid's
53/// type predicate (e.g. `wdt:P31`) instead of auto-detection. Uses the default
54/// [`PyramidAlgo::Louvain`] community algorithm — byte-identical to before.
55pub fn build_pyramid_meta_with(
56    dict: &Dictionary,
57    triples: &[(u32, u32, u32)],
58    budget: usize,
59    type_override: Option<&str>,
60) -> (Vec<u8>, u16) {
61    build_pyramid_meta_algo(dict, triples, budget, type_override, PyramidAlgo::Louvain)
62}
63
64/// Like [`build_pyramid_meta_with`], but selects the community [`PyramidAlgo`].
65/// [`PyramidAlgo::Types`] partitions by `rdf:type` — the deterministic,
66/// parallelizable alternative to Louvain (one linear pass, no modularity) that
67/// still emits the full summary + `query_stats`; it falls back to Louvain when the
68/// graph has no usable typing. Everything downstream of the dendrogram (round
69/// choice, summary, schema pyramid, planner stats) is shared across algorithms.
70pub fn build_pyramid_meta_algo(
71    dict: &Dictionary,
72    triples: &[(u32, u32, u32)],
73    budget: usize,
74    type_override: Option<&str>,
75    algo: PyramidAlgo,
76) -> (Vec<u8>, u16) {
77    build_pyramid_meta_layout(
78        dict,
79        triples,
80        budget,
81        type_override,
82        algo,
83        PyramidLayout::Combined,
84    )
85}
86
87/// What a pyramid-meta section carries besides the community summary.
88#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
89pub enum PyramidLayout {
90    /// Everything, as every build up to 0.3.x wrote it: the summary, the
91    /// per-community descriptors, the schema pyramid, `query_stats`, the
92    /// characteristic sets and the bounded (8,192-entry) label block.
93    #[default]
94    Combined,
95    /// Only what needs the community partition: the summary and the
96    /// per-community descriptors. The schema, `query_stats` and the
97    /// characteristic sets live in the SCHEMA section (kind 9) and the labels
98    /// in the LABEL_INDEX section (kind 8) — what `rete build` writes.
99    CommunityOnly,
100}
101
102/// Like [`build_pyramid_meta_algo`], choosing the [`PyramidLayout`].
103pub fn build_pyramid_meta_layout(
104    dict: &Dictionary,
105    triples: &[(u32, u32, u32)],
106    budget: usize,
107    type_override: Option<&str>,
108    algo: PyramidAlgo,
109    layout: PyramidLayout,
110) -> (Vec<u8>, u16) {
111    // Optional sub-phase timing (set RETE_BUILD_TIMING=1) — the pyramid build is
112    // the dominant cost of a big `rete build`; this shows where inside it.
113    // `Instant::now()` must stay behind the flag: `std::time` is unsupported on
114    // `wasm32-unknown-unknown` and panics ("time not implemented"), so an
115    // unconditional clock read would break every in-browser `build()`.
116    let timing = std::env::var_os("RETE_BUILD_TIMING").is_some();
117    let mut t = timing.then(std::time::Instant::now);
118    let mut lap = |label: &str| {
119        if let Some(t0) = &mut t {
120            eprintln!(
121                "  [pyramid] {label}: {:.0} ms",
122                t0.elapsed().as_secs_f64() * 1000.0
123            );
124            *t0 = std::time::Instant::now();
125        }
126    };
127
128    // The community partition — the only step that differs by algorithm.
129    let louvain = |lap: &mut dyn FnMut(&str)| {
130        let g = project_graph(dict, triples);
131        lap("project_graph");
132        let d = build_dendrogram(&g);
133        lap("build_dendrogram (Louvain)");
134        d
135    };
136    let dend = match algo {
137        PyramidAlgo::Louvain => louvain(&mut lap),
138        PyramidAlgo::Types => {
139            match crate::schema_pyramid::build_type_dendrogram(dict, triples, type_override) {
140                Some(d) => {
141                    lap("build_type_dendrogram");
142                    d
143                }
144                None => {
145                    eprintln!(
146                        "  [pyramid] --pyramid-algo types: no usable rdf:type \
147                         predicate — falling back to louvain"
148                    );
149                    louvain(&mut lap)
150                }
151            }
152        }
153    };
154    let round = choose_round_for_budget(dict, triples, &dend, budget);
155    lap("choose_round_for_budget");
156    let summary = summarize(dict, triples, &dend, round);
157    lap("summarize");
158    // Attach the v2 schema pyramid (the non-exclusive subClassOf DAG + per-level
159    // type rollups + per-level lateral class relations + per-community
160    // descriptors). Empty when the graph has no usable typing, in which case the
161    // encoding stays byte-identical to a v1 pyramid-meta.
162    if layout == PyramidLayout::CommunityOnly {
163        // Descriptors only: the schema pyramid's other fields, the planner
164        // stats, the shapes and the labels have sections of their own.
165        let descriptors = crate::schema_pyramid::build_community_descriptors(
166            dict,
167            triples,
168            &dend,
169            round,
170            type_override,
171        );
172        lap("build_community_descriptors");
173        let meta = PyramidMeta::new(round as u32, summary, &[]).with_schema(
174            Vec::new(),
175            Vec::new(),
176            Vec::new(),
177            descriptors,
178            Vec::new(),
179            Vec::new(),
180            Vec::new(),
181        );
182        let out = (meta.encode(), dend.rounds() as u16);
183        lap("encode");
184        return out;
185    }
186    let sp = crate::schema_pyramid::build_schema_pyramid_with(
187        dict,
188        triples,
189        &dend,
190        round,
191        type_override,
192    );
193    lap("build_schema_pyramid");
194    let predicate_stats = compute_predicate_stats(triples);
195    lap("compute_predicate_stats");
196    let char_sets = compute_char_sets(triples);
197    lap("compute_char_sets");
198    let label_index = compute_label_index(dict, triples);
199    lap("compute_label_index");
200    let meta = PyramidMeta::new(round as u32, summary, &[])
201        .with_schema(
202            sp.class_hierarchy,
203            sp.level_rollups,
204            sp.level_links,
205            sp.descriptors,
206            sp.subclass_cycles,
207            sp.disjoint_pairs,
208            sp.equivalent_pairs,
209        )
210        .with_predicate_stats(predicate_stats)
211        .with_char_sets(char_sets)
212        .with_label_index(label_index);
213    let out = (meta.encode(), dend.rounds() as u16);
214    lap("encode");
215    out
216}
217
218/// The label predicates a [`compute_label_index`] entry can come from — the
219/// common "human-readable name of this subject" terms, angle-bracketed as the
220/// dictionary stores them. Order is irrelevant (we union their ids). The same
221/// list the LABEL_INDEX section's rule L1 uses.
222const LABEL_PREDICATES: &[&str] = crate::label_index::DIRECT_LABEL_PREDICATES;
223
224/// Build the bounded, **legacy** label block of a [`PyramidLayout::Combined`]
225/// pyramid-meta (the LABEL_INDEX section, kind 8, replaces it): the display labels of
226/// the most-connected labeled subjects, sorted by the label's lowercased form.
227/// Ranking keeps autocomplete useful on a huge graph (the prominent entities
228/// survive the bound); a graph with fewer than `MAX_LABELS` labels keeps them
229/// all. Deterministic (degree, then subject id, then label) so builds are
230/// reproducible. O(triples) transient memory, freed before the file is written.
231fn compute_label_index(
232    dict: &Dictionary,
233    triples: &[(u32, u32, u32)],
234) -> Vec<crate::meta::LabelEntry> {
235    use crate::terms::{is_literal, literal_lexical};
236    use std::collections::{HashMap, HashSet};
237    const MAX_LABELS: usize = 8192;
238
239    // Resolve the label predicates that actually occur in this graph.
240    let label_pids: HashSet<u32> = LABEL_PREDICATES
241        .iter()
242        .filter_map(|p| dict.predicate_id(p))
243        .collect();
244    if label_pids.is_empty() {
245        return Vec::new();
246    }
247    // Subject degree (triple count) — the ranking used to bound the index.
248    let mut degree: HashMap<u32, u32> = HashMap::new();
249    for &(s, _p, _o) in triples {
250        *degree.entry(s).or_insert(0) += 1;
251    }
252    // Candidate (subject, label) pairs, deduped on (subject, lowercased label).
253    let mut seen: HashSet<(u32, String)> = HashSet::new();
254    let mut candidates: Vec<(u32, String, u32)> = Vec::new(); // (degree, label, subject)
255    for &(s, p, o) in triples {
256        if !label_pids.contains(&p) {
257            continue;
258        }
259        let Some(term) = dict.object_term(o) else {
260            continue;
261        };
262        if !is_literal(&term) {
263            continue;
264        }
265        let Some(label) = literal_lexical(&term) else {
266            continue;
267        };
268        if label.is_empty() {
269            continue;
270        }
271        if seen.insert((s, label.to_lowercase())) {
272            candidates.push((*degree.get(&s).unwrap_or(&0), label, s));
273        }
274    }
275    // Keep the most-connected entities when over budget: rank by degree desc,
276    // then subject asc, then label asc (deterministic).
277    if candidates.len() > MAX_LABELS {
278        candidates.sort_by(|a, b| {
279            b.0.cmp(&a.0)
280                .then_with(|| a.2.cmp(&b.2))
281                .then_with(|| a.1.cmp(&b.1))
282        });
283        candidates.truncate(MAX_LABELS);
284    }
285    // Final order: by lowercased label (search key), then label, then subject.
286    candidates.sort_by(|a, b| {
287        a.1.to_lowercase()
288            .cmp(&b.1.to_lowercase())
289            .then_with(|| a.1.cmp(&b.1))
290            .then_with(|| a.2.cmp(&b.2))
291    });
292    candidates
293        .into_iter()
294        .map(|(_deg, label, subject)| crate::meta::LabelEntry { label, subject })
295        .collect()
296}
297
298/// Build the default-scope full-text index section (`token → subjects`) over
299/// every literal object of `triples`: tokenize each literal into words and
300/// record the subject that carries it. Empty when the graph has no literals.
301/// This is the pre-0.4.0 index, byte for byte; the builds call
302/// `text_index::build_section` with their scope (`--text-index-graphs`,
303/// `--text-index-include`/`-exclude`). O(literal bytes) transient. The token
304/// table is compressed with `writer_codec` (the reader decompresses with
305/// `block_codec`).
306pub fn compute_text_index(dict: &Dictionary, triples: &[(u32, u32, u32)]) -> Vec<u8> {
307    crate::text_index::build_section(
308        dict,
309        &crate::text_index::TextIndexScope::default(),
310        triples,
311        std::iter::empty(),
312        writer_codec(),
313    )
314}
315
316/// The top entity **shapes** (characteristic sets): group subjects by the exact
317/// set of predicates they carry, keep the most common. Bounded to `MAX_CHAR_SETS`
318/// and sorted deterministically (by subject count, then predicate list) so the
319/// encoding is reproducible. O(triples) transient memory.
320fn compute_char_sets(triples: &[(u32, u32, u32)]) -> Vec<crate::meta::CharSet> {
321    use std::collections::{BTreeSet, HashMap};
322    const MAX_CHAR_SETS: usize = 128;
323    let mut by_subject: HashMap<u32, BTreeSet<u32>> = HashMap::new();
324    for &(s, p, _o) in triples {
325        by_subject.entry(s).or_default().insert(p);
326    }
327    let mut shapes: HashMap<Vec<u32>, u64> = HashMap::new();
328    for set in by_subject.into_values() {
329        *shapes.entry(set.into_iter().collect()).or_insert(0) += 1;
330    }
331    let mut v: Vec<crate::meta::CharSet> = shapes
332        .into_iter()
333        .map(|(predicates, subjects)| crate::meta::CharSet {
334            predicates,
335            subjects,
336        })
337        .collect();
338    v.sort_by(|a, b| {
339        b.subjects
340            .cmp(&a.subjects)
341            .then_with(|| a.predicates.cmp(&b.predicates))
342    });
343    v.truncate(MAX_CHAR_SETS);
344    v
345}
346
347/// Per-predicate cardinality for the cost-based planner, in one pass over the
348/// triples (deduped, so a per-(subject,predicate) count is its distinct-object
349/// count). Returned sorted by predicate id for a reproducible encoding. Holds a
350/// transient `(subject -> count, object -> count)` map per predicate — O(triples)
351/// memory, freed before the file is written.
352fn compute_predicate_stats(triples: &[(u32, u32, u32)]) -> Vec<crate::meta::PredStat> {
353    use std::collections::HashMap;
354    #[allow(clippy::type_complexity)]
355    let mut acc: HashMap<u32, (HashMap<u32, u32>, HashMap<u32, u32>, u64)> = HashMap::new();
356    for &(s, p, o) in triples {
357        let e = acc.entry(p).or_default();
358        *e.0.entry(s).or_insert(0) += 1;
359        *e.1.entry(o).or_insert(0) += 1;
360        e.2 += 1;
361    }
362    let mut stats: Vec<crate::meta::PredStat> = acc
363        .into_iter()
364        .map(|(predicate, (subj, obj, count))| crate::meta::PredStat {
365            predicate,
366            count,
367            distinct_subjects: subj.len() as u64,
368            distinct_objects: obj.len() as u64,
369            max_objects_per_subject: subj.values().copied().max().unwrap_or(0),
370            max_subjects_per_object: obj.values().copied().max().unwrap_or(0),
371        })
372        .collect();
373    stats.sort_by_key(|p| p.predicate);
374    stats
375}
376
377/// No compression.
378pub const CODEC_NONE: u8 = 0;
379/// zstd compression (per section).
380pub const CODEC_ZSTD: u8 = 1;
381
382#[derive(Debug, thiserror::Error)]
383#[non_exhaustive]
384pub enum FileError {
385    #[error("header: {0}")]
386    Header(#[from] crate::header::HeaderError),
387    #[error("malformed container: {0}")]
388    Container(&'static str),
389    #[error("unknown codec: {0}")]
390    UnknownCodec(u8),
391    #[error("decompression failed: {0}")]
392    Decompress(std::io::Error),
393    #[error("io: {0}")]
394    Io(#[from] std::io::Error),
395}
396
397/// The codec the writer uses: zstd when the `compression` feature is on, else
398/// none. Reading honors whatever codec the header records (when supported).
399pub fn writer_codec() -> u8 {
400    if cfg!(feature = "compression") {
401        CODEC_ZSTD
402    } else {
403        CODEC_NONE
404    }
405}
406
407/// Intersection of two ascending-sorted, deduped id lists — the AND of two
408/// posting lists in a multi-word text search. Linear merge, output sorted.
409fn intersect_sorted(a: &[u32], b: &[u32]) -> Vec<u32> {
410    let mut out = Vec::with_capacity(a.len().min(b.len()));
411    let (mut i, mut j) = (0, 0);
412    while i < a.len() && j < b.len() {
413        match a[i].cmp(&b[j]) {
414            std::cmp::Ordering::Less => i += 1,
415            std::cmp::Ordering::Greater => j += 1,
416            std::cmp::Ordering::Equal => {
417                out.push(a[i]);
418                i += 1;
419                j += 1;
420            }
421        }
422    }
423    out
424}
425
426pub(crate) fn compress(codec: u8, bytes: &[u8]) -> Vec<u8> {
427    match codec {
428        // Level and frame shape are writer knobs (`crate::write_opts`); any
429        // combination decodes with every conformant zstd decoder.
430        #[cfg(feature = "compression")]
431        CODEC_ZSTD => {
432            let level = crate::write_opts::zstd_level();
433            match crate::write_opts::frame_shape() {
434                crate::write_opts::FrameShape::Stream => zstd::encode_all(bytes, level),
435                crate::write_opts::FrameShape::Sized => zstd::bulk::compress(bytes, level),
436            }
437            .expect("zstd encode is infallible in-memory")
438        }
439        _ => bytes.to_vec(),
440    }
441}
442
443/// The streaming twin of [`compress`]: the same codec, level and frame shape,
444/// read from `src` (exactly `src_len` bytes) and written to `dst`, for a
445/// payload the memory-bounded build keeps on disk (the TEXT_INDEX token
446/// table). A streaming frame is what `zstd::encode_all` writes, through the
447/// same `Encoder`. A sized frame is told the source size up front
448/// (`set_pledged_src_size`), so it picks the same small-input parameters and
449/// records the same content size as `zstd::bulk::compress`, without holding
450/// the payload. The encoder cuts blocks from its own input buffer, not from
451/// how the input was handed to it, so how `src` hands out its bytes never
452/// changes the output.
453///
454/// It equals [`compress`]'s one-shot output only while the payload fits the
455/// level's window (4 MiB at level 9): past it the streaming encoder's ring
456/// buffer wraps and it finds slightly different matches. A payload that one
457/// writer streams and another holds whole must therefore go through
458/// [`compress_streamed`] on the in-memory side — the token table does.
459/// `text_index_spill` pins both facts.
460#[cfg(not(target_arch = "wasm32"))]
461pub(crate) fn compress_to(
462    codec: u8,
463    src_len: u64,
464    src: &mut dyn std::io::Read,
465    dst: &mut dyn std::io::Write,
466) -> std::io::Result<()> {
467    match codec {
468        #[cfg(feature = "compression")]
469        CODEC_ZSTD => zstd_stream(src_len, src, dst),
470        _ => {
471            let _ = src_len;
472            std::io::copy(src, dst).map(|_| ())
473        }
474    }
475}
476
477/// [`compress`] through the streaming encoder: byte for byte what
478/// [`compress_to`] writes for the same payload, whatever its size. For the
479/// in-memory writer of a payload the memory-bounded build compresses from
480/// disk (the TEXT_INDEX token table), so the two builds write the same bytes
481/// past the zstd window too. Below it this is [`compress`]'s output.
482pub(crate) fn compress_streamed(codec: u8, bytes: &[u8]) -> Vec<u8> {
483    match codec {
484        #[cfg(feature = "compression")]
485        CODEC_ZSTD => {
486            let mut out = Vec::with_capacity(bytes.len() / 4 + 64);
487            zstd_stream(bytes.len() as u64, &mut &bytes[..], &mut out)
488                .expect("zstd encode is infallible in-memory");
489            out
490        }
491        _ => bytes.to_vec(),
492    }
493}
494
495/// The streaming zstd encoder at the writer's level and frame shape.
496#[cfg(feature = "compression")]
497fn zstd_stream(
498    src_len: u64,
499    src: &mut dyn std::io::Read,
500    dst: &mut dyn std::io::Write,
501) -> std::io::Result<()> {
502    let mut enc = zstd::stream::Encoder::new(dst, crate::write_opts::zstd_level())?;
503    if crate::write_opts::frame_shape() == crate::write_opts::FrameShape::Sized {
504        enc.set_pledged_src_size(Some(src_len))?;
505        enc.include_contentsize(true)?;
506    }
507    let copied = std::io::copy(src, &mut enc)?;
508    if copied != src_len {
509        return Err(std::io::Error::other(format!(
510            "zstd stream: {copied} bytes read, {src_len} announced"
511        )));
512    }
513    enc.finish()?;
514    Ok(())
515}
516
517pub(crate) fn decompress(codec: u8, bytes: &[u8]) -> Result<Vec<u8>, FileError> {
518    match codec {
519        CODEC_NONE => Ok(bytes.to_vec()),
520        // Native (the `compression` feature — always on for the CLI, off for
521        // wasm) decodes through the C `zstd` library, the same one
522        // [`compress`] encodes with. It is roughly 2× the pure-Rust decoder
523        // and speeds every read path, not just export; the multi-pass tight
524        // budget of a bounded export (`--memory-budget-mb`) re-decodes the
525        // dictionary, so a faster decoder is the single largest lever there.
526        // A valid zstd frame decodes to the same bytes under any conformant
527        // decoder, so the output is byte-for-byte what `ruzstd` produced.
528        #[cfg(feature = "compression")]
529        CODEC_ZSTD => zstd::decode_all(bytes).map_err(FileError::Decompress),
530        // Pure-Rust decode so a target WITHOUT a C toolchain (wasm, or a
531        // `--no-default-features` build) can still read compressed files.
532        // This is the ONLY decode path wasm ever compiles — the C `zstd`
533        // crate is gated behind `compression`, which wasm never enables.
534        #[cfg(not(feature = "compression"))]
535        CODEC_ZSTD => {
536            use std::io::Read;
537            let mut dec = ruzstd::decoding::StreamingDecoder::new(bytes)
538                .map_err(|e| FileError::Decompress(std::io::Error::other(e.to_string())))?;
539            let mut out = Vec::new();
540            dec.read_to_end(&mut out).map_err(FileError::Decompress)?;
541            Ok(out)
542        }
543        other => Err(FileError::UnknownCodec(other)),
544    }
545}
546
547/// The `Frame_Content_Size` a zstd frame header declares (RFC 8878 §3.1.1.1),
548/// or `None` when the frame does not state it (or `bytes` is not a zstd
549/// frame). Reads the header only.
550pub(crate) fn zstd_frame_content_size(b: &[u8]) -> Option<u64> {
551    if b.len() < 5 || b[..4] != [0x28, 0xB5, 0x2F, 0xFD] {
552        return None;
553    }
554    let fhd = b[4];
555    let fcs_flag = fhd >> 6;
556    let single_segment = fhd & 0x20 != 0;
557    let did_size = [0usize, 1, 2, 4][(fhd & 0x03) as usize];
558    let mut pos = 5 + usize::from(!single_segment) + did_size;
559    let fcs_size = match (fcs_flag, single_segment) {
560        (0, false) => return None,
561        (0, true) => 1,
562        (1, _) => 2,
563        (2, _) => 4,
564        _ => 8,
565    };
566    let f = b.get(pos..pos + fcs_size)?;
567    pos += fcs_size;
568    let _ = pos;
569    let mut v = 0u64;
570    for (i, x) in f.iter().enumerate() {
571        v |= (*x as u64) << (8 * i);
572    }
573    Some(if fcs_size == 2 { v + 256 } else { v })
574}
575
576/// [`decompress`] with the output bounded by `max` bytes: a frame whose header
577/// declares more is refused before any decoding, and a frame that does not
578/// declare its size is decoded at most `max + 1` bytes far and refused past
579/// `max`. Format `0x06` records each section's largest block, so a corrupt or
580/// hostile frame cannot make a reader allocate more than the writer did.
581pub(crate) fn decompress_bounded(codec: u8, bytes: &[u8], max: u64) -> Result<Vec<u8>, FileError> {
582    let too_big = |n: u64| {
583        FileError::Decompress(std::io::Error::other(format!(
584            "a block of {n} bytes exceeds its section's {max}-byte maximum"
585        )))
586    };
587    match codec {
588        CODEC_NONE => {
589            if bytes.len() as u64 > max {
590                return Err(too_big(bytes.len() as u64));
591            }
592            Ok(bytes.to_vec())
593        }
594        CODEC_ZSTD => {
595            let declared = zstd_frame_content_size(bytes);
596            if let Some(n) = declared.filter(|&n| n > max) {
597                return Err(too_big(n));
598            }
599            #[cfg(feature = "compression")]
600            {
601                if let Some(n) = declared {
602                    // One-shot into exactly the declared size: a frame that
603                    // lies about it fails here instead of growing the buffer.
604                    return zstd::bulk::decompress(bytes, n as usize)
605                        .map_err(FileError::Decompress);
606                }
607                use std::io::Read;
608                let mut out = Vec::new();
609                let mut dec = zstd::stream::read::Decoder::with_buffer(bytes)
610                    .map_err(FileError::Decompress)?;
611                dec.window_log_max(27).map_err(FileError::Decompress)?;
612                dec.take(max.saturating_add(1))
613                    .read_to_end(&mut out)
614                    .map_err(FileError::Decompress)?;
615                if out.len() as u64 > max {
616                    return Err(too_big(out.len() as u64));
617                }
618                Ok(out)
619            }
620            #[cfg(not(feature = "compression"))]
621            {
622                use std::io::Read;
623                let dec = ruzstd::decoding::StreamingDecoder::new(bytes)
624                    .map_err(|e| FileError::Decompress(std::io::Error::other(e.to_string())))?;
625                let mut out = Vec::new();
626                dec.take(max.saturating_add(1))
627                    .read_to_end(&mut out)
628                    .map_err(FileError::Decompress)?;
629                if out.len() as u64 > max {
630                    return Err(too_big(out.len() as u64));
631                }
632                Ok(out)
633            }
634        }
635        other => Err(FileError::UnknownCodec(other)),
636    }
637}
638
639/// The decompressed length of one stored block — from its frame header when
640/// it states it, else by decoding it (bounded by `ceiling`). What the `0x06`
641/// writer records as a section's largest block.
642pub(crate) fn block_raw_len(codec: u8, bytes: &[u8], ceiling: u64) -> Result<u64, FileError> {
643    match codec {
644        CODEC_NONE => Ok(bytes.len() as u64),
645        CODEC_ZSTD => match zstd_frame_content_size(bytes) {
646            Some(n) => Ok(n),
647            None => Ok(decompress_bounded(codec, bytes, ceiling)?.len() as u64),
648        },
649        other => Err(FileError::UnknownCodec(other)),
650    }
651}
652
653/// Bytes this close are cheaper fetched as one read than as two round trips:
654/// tiles are laid back-to-back so this only ever bridges a tile already made
655/// resident by an earlier window — keep it tight to avoid re-fetching it.
656const TILE_COALESCE_GAP: u64 = 4096;
657
658/// The dictionary chunks a query's output terms touch are scattered across the
659/// section (terms are sorted, output ids are not), so byte-adjacency is rare.
660/// A wider gap trades a little over-fetch for far fewer round trips — the right
661/// call on a latency-bound remote read, where one skipped 64 KiB chunk is much
662/// cheaper than another request's RTT.
663const DICT_COALESCE_GAP: u64 = 64 * 1024;
664
665/// The default `rete export --memory-budget-mb` when the flag is omitted — the
666/// same number `rete build`/`merge` default to, so one figure governs both the
667/// build and the export of a file. Comfortably above a sub-4-GB-dictionary
668/// file's decompressed dictionary (those export in a single pass at today's
669/// speed), and far under a large VM's RAM so a 50-GB-class file is bounded
670/// rather than OOM. See [`split_memory_budget`].
671pub const DEFAULT_EXPORT_BUDGET_MB: u64 = 4096;
672
673/// The largest default memory budget, in MiB, of a lazily-opened file — the
674/// bound on its dictionary chunk cache, index tile cache and raw block cache,
675/// split as [`split_memory_budget`] describes. The default actually applied is
676/// sized from the memory the process can still use, half of it, between
677/// [`crate::sysmem::MIN_DEFAULT_QUERY_BUDGET_MB`] and this figure (see
678/// [`crate::sysmem::default_query_budget_mb`]; this figure whenever the memory
679/// cannot be read). `rete sparql --memory-budget-mb` (and the other query
680/// commands) default to it too, and so does every client that opens through
681/// [`Rete::open_ranged_lazy`] (Python, R, Java, the relay), so a long-lived
682/// session no longer grows with every tile a query ever touched. The same
683/// figure as [`DEFAULT_EXPORT_BUDGET_MB`].
684pub const DEFAULT_QUERY_BUDGET_MB: u64 = DEFAULT_EXPORT_BUDGET_MB;
685
686/// The [`Rete::spill_threshold`] a memory budget implies: an eighth of it. A
687/// spilling `COUNT(DISTINCT …)` buffer can briefly occupy about twice its
688/// threshold while it sorts, so this keeps it within a quarter of the budget.
689pub fn spill_threshold_for(budget: u64) -> u64 {
690    budget / 8
691}
692
693/// The lowest byte cap the dictionary chunk cache is ever given, whatever the
694/// budget. It only has to guarantee forward progress and hold a handful of
695/// chunks (a single front-coded run can dwarf the 64 KiB chunk budget — a
696/// base64-embedded image, say). Correctness never depends on it: the cache's
697/// `Arc` handout keeps a chunk alive across its one decode even if it is
698/// evicted the instant it is inserted, so a full dump completes at *any* cap.
699/// This floor only keeps a pathologically small `--memory-budget-mb` from
700/// thrashing the whole dictionary chunk-by-chunk.
701const MIN_DICT_CACHE_CAP: u64 = 64 << 20; // 64 MiB
702
703/// Upper bound of the budget share [`split_memory_budget`] holds back for
704/// transient, uncached allocations (see the `reserve` there).
705const TRANSIENT_RESERVE_MAX: u64 = 128 << 20; // 128 MiB
706
707/// How a total `--memory-budget-mb` splits across the export's three bounded
708/// consumers. The three caps sum to the total less a small reserve (≤ 128 MiB)
709/// for the window's buffers and one window's in-flight reads, which no cache
710/// owns. [`Rete::set_memory_budget`] further subtracts the open's fixed
711/// directories ([`Rete::fixed_overhead`]) from the dictionary share, so the
712/// live heap of a dump stays within the total.
713#[derive(Debug, Clone, Copy)]
714pub struct MemoryBudget {
715    /// Raw file-block LRU (the `BlockCacheReader` that backs the ranged reader).
716    pub block_cache: u64,
717    /// Decompressed dictionary chunk bodies (the old ~4.3 GB export floor).
718    pub dict_cache: u64,
719    /// Decoded index tiles of the scanned permutation.
720    pub tile_cache: u64,
721}
722
723/// Split a total byte budget across the block cache, the dictionary chunk cache
724/// and the index tile cache (bounded-export phase 2, plan §5.2). `u64::MAX`
725/// (unlimited, e.g. `--in-memory` or `--memory-budget-mb 0`) passes straight
726/// through to every consumer.
727///
728/// The dictionary is the dominant consumer on scholar files (literal-heavy
729/// object section), and its chunks recur across windows, so it gets the bulk of
730/// the budget; the tile cache is small (a scanned permutation is far smaller
731/// than the dictionary) and the block cache stays modest, capped like every
732/// other ranged reader at [`crate::block_cache::DEFAULT_CACHE_CAP`].
733pub fn split_memory_budget(total: u64) -> MemoryBudget {
734    if total == u64::MAX {
735        return MemoryBudget {
736            block_cache: u64::MAX,
737            dict_cache: u64::MAX,
738            tile_cache: u64::MAX,
739        };
740    }
741    // Raw file blocks: min(256 MiB, total/8). Repeated directory/header reads
742    // must stay cheap, but this cache is not where the floor lived.
743    let block_cache = (total / 8).min(crate::block_cache::DEFAULT_CACHE_CAP);
744    // Scanned-permutation tiles: modest — tens of MB on epfl-class files. Cap
745    // at total/16 (≤ 256 MiB) so the dictionary keeps the rest.
746    let tile_cache = (total / 16).min(256 << 20);
747    // Held back for what no cache owns: the window's id and string buffers,
748    // the coalesced read spans and freshly decoded bodies of one window's
749    // prefetch before they enter the cache, and the program itself. Measured at
750    // ~80-120 MB on epfl and openaire-relation (dev/export-memory), so 1/32 of
751    // the budget, capped at 128 MiB.
752    let reserve = (total / 32).min(TRANSIENT_RESERVE_MAX);
753    // The dictionary chunk cache gets everything else, floored so a tiny budget
754    // still holds a working set of chunks.
755    let dict_cache = total
756        .saturating_sub(block_cache)
757        .saturating_sub(tile_cache)
758        .saturating_sub(reserve)
759        .max(MIN_DICT_CACHE_CAP);
760    MemoryBudget {
761        block_cache,
762        dict_cache,
763        tile_cache,
764    }
765}
766
767/// Heap an open `Rete` holds outside its bounded caches, by structure — see
768/// [`Rete::fixed_overhead`].
769#[derive(Debug, Clone, Copy, Default)]
770pub struct FixedOverhead {
771    /// The four dictionary sections' chunk directories.
772    pub dict_directory: u64,
773    /// The default graph's tile directories (every carried permutation).
774    pub index_directory: u64,
775    /// The named-graph directory plus the open graphs' tile directories.
776    pub named_graphs: u64,
777}
778
779impl FixedOverhead {
780    /// The sum of every part.
781    pub fn total(&self) -> u64 {
782        self.dict_directory + self.index_directory + self.named_graphs
783    }
784}
785
786/// Format a cache byte cap for `RETE_OPEN_DEBUG` (`unlimited` for `u64::MAX`).
787fn fmt_cap(cap: u64) -> String {
788    if cap == u64::MAX {
789        "unlimited".to_string()
790    } else {
791        format!("{}MiB", cap >> 20)
792    }
793}
794
795/// Fetch a set of ascending, disjoint byte ranges, coalescing ranges whose gap
796/// is at most `gap` into one span, then fetching the spans through
797/// [`RangeReader::read_many`] (which a parallelizable reader issues
798/// concurrently). Returns each requested range's bytes in order; `None` if any
799/// read fails.
800pub(crate) fn read_coalesced<R: RangeReader + ?Sized>(
801    reader: &R,
802    ranges: &[ByteRange],
803    gap: u64,
804) -> Option<Vec<Vec<u8>>> {
805    coalesced(reader, ranges, gap, false)
806}
807
808/// [`read_coalesced`] for ranges known to be scattered (a probe window's
809/// tiles): the spans go through [`RangeReader::read_many_scattered`], so the
810/// reader asks for them as they are.
811pub(crate) fn read_coalesced_scattered<R: RangeReader + ?Sized>(
812    reader: &R,
813    ranges: &[ByteRange],
814    gap: u64,
815) -> Option<Vec<Vec<u8>>> {
816    coalesced(reader, ranges, gap, true)
817}
818
819fn coalesced<R: RangeReader + ?Sized>(
820    reader: &R,
821    ranges: &[ByteRange],
822    gap: u64,
823    scattered: bool,
824) -> Option<Vec<Vec<u8>>> {
825    // Build the coalesced spans and remember which span each input range maps
826    // into, so the fetched span blobs can be sliced back apart in order.
827    let mut spans: Vec<(u64, u64)> = Vec::new();
828    let mut span_of: Vec<usize> = Vec::with_capacity(ranges.len());
829    let mut i = 0;
830    while i < ranges.len() {
831        let start = ranges[i].offset;
832        let mut end = ranges[i].offset.checked_add(ranges[i].len)?;
833        let mut j = i + 1;
834        while j < ranges.len() {
835            let r = &ranges[j];
836            if r.offset < end || r.offset - end > gap {
837                break;
838            }
839            end = r.offset.checked_add(r.len)?;
840            j += 1;
841        }
842        let si = spans.len();
843        spans.push((start, end - start));
844        for _ in i..j {
845            span_of.push(si);
846        }
847        i = j;
848    }
849    let blobs = if scattered {
850        reader.read_many_scattered(&spans)
851    } else {
852        reader.read_many(&spans)
853    }
854    .ok()?;
855    if blobs.len() != spans.len() {
856        return None;
857    }
858    let mut out = Vec::with_capacity(ranges.len());
859    for (k, r) in ranges.iter().enumerate() {
860        let (span_start, _) = spans[span_of[k]];
861        let blob = &blobs[span_of[k]];
862        out.push(checked_slice(blob, r.offset - span_start, r.len)?.to_vec());
863    }
864    Some(out)
865}
866
867/// Content hash (first 16 bytes of blake3) over the file payload sections.
868/// Identifies the immutable content independent of the header.
869fn content_hash(parts: &[&[u8]]) -> [u8; 16] {
870    let mut h = blake3::Hasher::new();
871    for p in parts {
872        h.update(p);
873    }
874    let mut out = [0u8; 16];
875    out.copy_from_slice(&h.finalize().as_bytes()[..16]);
876    out
877}
878
879/// A resolved triple as terms.
880pub type TermTriple = (String, String, String);
881
882/// What a filtered dump will cost, before it runs — see
883/// [`Rete::dump_plan`](crate::Rete::dump_plan).
884#[derive(Debug, Clone, Copy, PartialEq, Eq)]
885pub struct DumpPlan {
886    /// The routed scan this dump will run. `None` when it provably matches
887    /// nothing without touching the index: an absent graph IRI, or a bound term
888    /// the dictionary does not contain.
889    pub scan: Option<crate::index::ScanPlan>,
890    /// The file's whole dictionary section — the **ceiling** on what resolving
891    /// the matched rows' terms can fault, and on a literal-heavy file the
892    /// dominant term in a dump's real cost.
893    ///
894    /// It is deliberately not an estimate. A dump that matches one row faults
895    /// only the chunks that row's three terms live in, but nothing in the tile
896    /// directories says which chunks those are or how big they are, and a
897    /// preview that guessed would be worse than one that states the bound.
898    pub dictionary_bytes: u64,
899}
900
901/// One labelled byte region of a `.rete` file image (see
902/// [`Rete::file_layout`]). `kind` is a stable machine tag: `header`,
903/// `metadata`, `dictionary`, `directory`, `tile`, `pyramid`, `named-graphs`.
904#[derive(Debug, Clone)]
905pub struct LayoutSegment {
906    pub kind: &'static str,
907    pub label: String,
908    pub offset: u64,
909    pub len: u64,
910}
911
912/// A byte range in the `.rete` file image.
913#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
914pub struct ByteRange {
915    pub offset: u64,
916    pub len: u64,
917}
918
919impl ByteRange {
920    pub fn end(self) -> u64 {
921        self.offset + self.len
922    }
923}
924
925/// Why a triple-pattern result is present in the file.
926#[derive(Debug, Clone, PartialEq, Eq)]
927pub struct TripleProvenance {
928    /// Matched triple resolved to canonical N-Triples tokens.
929    pub terms: TermTriple,
930    /// Matched triple in dictionary ID space.
931    pub ids: Triple,
932    /// Named graph IRI, or `None` for the default graph.
933    pub graph: Option<String>,
934    /// The resolved ID-space pattern that was matched.
935    pub matched_pattern: Pattern,
936    /// Permutation selected to answer the pattern.
937    pub index_permutation: IndexPermutation,
938    /// File byte range containing the dictionary container.
939    pub dictionary_range: ByteRange,
940    /// File byte range containing the permutation index container.
941    pub index_range: ByteRange,
942    /// File byte range containing the selected permutation payload inside the
943    /// index container.
944    pub index_section_range: ByteRange,
945    /// File byte range containing the pyramid metadata, when present.
946    pub pyramid_range: Option<ByteRange>,
947    /// Physical tile identifier, once tile directories are materialized.
948    /// Physical tile identifier (`PERM/index`, e.g. `POS/3`) for tiled (v0.2)
949    /// files; `None` for pre-tiling files.
950    pub tile: Option<String>,
951    /// File byte range of that (compressed) tile — the exact bytes a ranged
952    /// client would fetch to re-derive this match.
953    pub tile_range: Option<ByteRange>,
954}
955
956/// Encode a length-prefixed container of byte sections, each compressed with
957/// `codec` independently (so a range-reading client decompresses only the
958/// sections it fetches). Stored length is the *compressed* length.
959fn encode_container(sections: &[&[u8]], codec: u8) -> Vec<u8> {
960    let mut out = Vec::new();
961    write_uvarint(&mut out, sections.len() as u64);
962    for s in sections {
963        let payload = compress(codec, s);
964        write_uvarint(&mut out, payload.len() as u64);
965        out.extend_from_slice(&payload);
966    }
967    out
968}
969
970/// Decode a container into owned, decompressed sections.
971fn decode_container(bytes: &[u8], codec: u8) -> Result<Vec<Vec<u8>>, FileError> {
972    let (n, mut pos) = read_uvarint(bytes).ok_or(FileError::Container("truncated count"))?;
973    // `n` is untrusted; each section needs ≥1 byte, so cap the pre-allocation at
974    // the buffer length rather than trusting the count (avoids an OOM on a bogus
975    // header pointing at a small region).
976    let n = wire_usize(n, "section count too large")?;
977    let mut out = Vec::with_capacity(n.min(bytes.len()));
978    for _ in 0..n {
979        let (len, used) = read_uvarint(bytes.get(pos..).unwrap_or(&[]))
980            .ok_or(FileError::Container("truncated length"))?;
981        pos += used;
982        // Checked: a hostile length near u64::MAX used to overflow `pos + len`
983        // (a panic), and on wasm32 `len as usize` kept the low 32 bits.
984        let end = pos
985            .checked_add(wire_usize(len, "section length too large")?)
986            .filter(|&end| end <= bytes.len())
987            .ok_or(FileError::Container("section overruns buffer"))?;
988        out.push(decompress(codec, &bytes[pos..end])?);
989        pos = end;
990    }
991    Ok(out)
992}
993
994fn checked_end(off: u64, len: u64) -> Result<u64, FileError> {
995    off.checked_add(len)
996        .ok_or(FileError::Container("section range overflows"))
997}
998
999/// A length or count read from the file, as an in-memory `usize` — or a clean
1000/// error. `as usize` truncates on wasm32, a 32-bit target: a length of
1001/// `2^32 + 5` read 5 bytes there and decoding went on as if nothing were wrong.
1002fn wire_usize(v: u64, what: &'static str) -> Result<usize, FileError> {
1003    usize::try_from(v).map_err(|_| FileError::Container(what))
1004}
1005
1006/// The in-memory slice `[off, off + len)` of `bytes`, or `None` when the range
1007/// does not fit this target's `usize` or overruns the buffer.
1008fn checked_slice(bytes: &[u8], off: u64, len: u64) -> Option<&[u8]> {
1009    let start = usize::try_from(off).ok()?;
1010    let end = start.checked_add(usize::try_from(len).ok()?)?;
1011    bytes.get(start..end)
1012}
1013
1014/// Per-chunk budget for dictionary section bodies — same reasoning as
1015/// [`crate::index::INDEX_TILE_BUDGET`]: one fetch, one decompress per touch.
1016const DICT_CHUNK_BUDGET: usize = crate::write_opts::DEFAULT_DICT_CHUNK_BYTES;
1017
1018/// Encode one dictionary section as a chunked payload (format v0.2):
1019/// `[header_len, raw header (term_count/interval/restart table)]
1020///  [num_chunks; per chunk: Δfirst_run, key_len, key, comp_len]
1021///  [individually compressed run-aligned body slices]`.
1022/// The header keeps its original encoding, so restart offsets stay valid in
1023/// the section's coordinate space.
1024///
1025/// `key` is the chunk's **routing separator**, not its first term: the
1026/// shortest string strictly above the previous chunk's last term and at most
1027/// the chunk's own first term (empty for chunk 0). See
1028/// [`crate::dict::SectionChunk::key`] for the invariant and
1029/// [`crate::dict::shortest_separator`] for the choice. Storing the first term
1030/// verbatim — what this wrote before — is the degenerate case of the same
1031/// invariant, so readers of either vintage route both correctly; the
1032/// separator is simply orders of magnitude smaller when the boundary term is a
1033/// long literal.
1034fn encode_chunked_dict_section(raw: &[u8], codec: u8) -> Vec<u8> {
1035    let meta = crate::dict::parse_meta(raw).unwrap_or(crate::dict::SectionMeta {
1036        term_count: 0,
1037        restart_interval: 1,
1038        restart_offsets: Vec::new(),
1039    });
1040    let body_start = meta
1041        .restart_offsets
1042        .first()
1043        .copied()
1044        .unwrap_or(raw.len() as u64);
1045    let header = &raw[..(body_start.min(raw.len() as u64)) as usize];
1046
1047    // Split runs into chunks by body-byte budget (whole runs only).
1048    let chunk_budget = crate::write_opts::dict_chunk_budget() as u64;
1049    let n_runs = meta.restart_offsets.len();
1050    let mut bounds: Vec<(usize, u64, u64)> = Vec::new(); // (first_run, start, end)
1051    let mut r = 0;
1052    while r < n_runs {
1053        let start = meta.restart_offsets[r];
1054        let mut r2 = r + 1;
1055        while r2 < n_runs && meta.restart_offsets[r2] - start < chunk_budget {
1056            r2 += 1;
1057        }
1058        let end = if r2 < n_runs {
1059            meta.restart_offsets[r2]
1060        } else {
1061            raw.len() as u64
1062        };
1063        bounds.push((r, start, end));
1064        r = r2;
1065    }
1066
1067    // Chunks are independent: compress them across all cores, as the index
1068    // tiles are (`par_iter` keeps the order, so the bytes are unchanged). At a
1069    // high `--zstd-level` this serial loop was most of a build's wall time.
1070    #[cfg(feature = "parallel")]
1071    let compressed: Vec<Vec<u8>> = {
1072        use rayon::prelude::*;
1073        bounds
1074            .par_iter()
1075            .map(|&(_, s, e)| compress(codec, &raw[s as usize..e as usize]))
1076            .collect()
1077    };
1078    #[cfg(not(feature = "parallel"))]
1079    let compressed: Vec<Vec<u8>> = bounds
1080        .iter()
1081        .map(|&(_, s, e)| compress(codec, &raw[s as usize..e as usize]))
1082        .collect();
1083    let mut out = Vec::new();
1084    write_uvarint(&mut out, header.len() as u64);
1085    out.extend_from_slice(header);
1086    write_uvarint(&mut out, bounds.len() as u64);
1087    let mut prev_run = 0usize;
1088    // The previous chunk's last term — the separator's lower bound. Decoding it
1089    // walks at most `restart_interval` front-coded entries of one chunk, which
1090    // is nothing beside the zstd pass over the same 64 KiB.
1091    let mut prev_last: Option<Vec<u8>> = None;
1092    for (i, (&(first_run, start, end), comp)) in bounds.iter().zip(&compressed).enumerate() {
1093        let key = if i == 0 {
1094            // Everything routes to chunk 0; it needs no separator at all.
1095            Vec::new()
1096        } else {
1097            let first = crate::dict::run_first_term(raw, start as usize).unwrap_or_default();
1098            match &prev_last {
1099                // Undecodable predecessor: the verbatim first term is always a
1100                // valid separator, just not a short one.
1101                None => first,
1102                Some(pl) => crate::dict::shortest_separator(pl, &first),
1103            }
1104        };
1105        if i + 1 < bounds.len() {
1106            // The last run of THIS chunk starts one run before the next chunk's.
1107            let last_run_off = meta.restart_offsets[bounds[i + 1].0 - 1] as usize;
1108            prev_last = crate::dict::run_last_term(raw, last_run_off, end as usize);
1109        }
1110        write_uvarint(&mut out, (first_run - prev_run) as u64);
1111        write_uvarint(&mut out, key.len() as u64);
1112        out.extend_from_slice(&key);
1113        write_uvarint(&mut out, comp.len() as u64);
1114        prev_run = first_run;
1115    }
1116    for comp in &compressed {
1117        out.extend_from_slice(comp);
1118    }
1119    out
1120}
1121
1122/// A parsed chunked-dict-section directory entry: the chunk's run/key/body
1123/// coordinates plus its compressed byte range *within the payload*. `key` is a
1124/// routing separator, not a term — see [`crate::dict::SectionChunk::key`].
1125pub(crate) struct DictChunkEntry {
1126    pub(crate) first_run: usize,
1127    pub(crate) key: Vec<u8>,
1128    body_start: u64,
1129    pub(crate) start: u64,
1130    pub(crate) end: u64,
1131}
1132
1133/// Parse a chunked dictionary section's header + directory (not the chunks).
1134/// `bytes` may be a prefix of the payload; compressed ranges validate against
1135/// `total_len`.
1136fn parse_chunked_dict_dir(
1137    bytes: &[u8],
1138    total_len: u64,
1139) -> Result<(crate::dict::SectionMeta, Vec<DictChunkEntry>), FileError> {
1140    let mut pos = 0usize;
1141    let take = |pos: &mut usize| -> Result<u64, FileError> {
1142        let (v, n) = read_uvarint(bytes.get(*pos..).unwrap_or(&[]))
1143            .ok_or(FileError::Container("truncated dict chunk directory"))?;
1144        *pos += n;
1145        Ok(v)
1146    };
1147    let header_len = wire_usize(take(&mut pos)?, "dict header too large")?;
1148    let header = bytes
1149        .get(pos..pos.saturating_add(header_len))
1150        .ok_or(FileError::Container("truncated dict header"))?;
1151    let meta = crate::dict::parse_meta(header)
1152        .map_err(|_| FileError::Container("malformed dict header"))?;
1153    pos += header_len;
1154
1155    let num_chunks = wire_usize(take(&mut pos)?, "dict chunk count too large")?;
1156    let mut entries = Vec::with_capacity(num_chunks.min(bytes.len()));
1157    let mut lens = Vec::with_capacity(num_chunks.min(bytes.len()));
1158    let mut prev_run = 0usize;
1159    for i in 0..num_chunks {
1160        let drun = wire_usize(take(&mut pos)?, "dict chunk run too large")?;
1161        let klen = wire_usize(take(&mut pos)?, "dict chunk key too large")?;
1162        let key = bytes
1163            .get(pos..pos.saturating_add(klen))
1164            .ok_or(FileError::Container("truncated dict chunk key"))?
1165            .to_vec();
1166        pos += klen;
1167        let clen = take(&mut pos)?;
1168        let first_run = next_chunk_run(i, prev_run, drun)?;
1169        let body_start = meta
1170            .restart_offsets
1171            .get(first_run)
1172            .copied()
1173            .ok_or(FileError::Container("dict chunk run out of range"))?;
1174        entries.push(DictChunkEntry {
1175            first_run,
1176            key,
1177            body_start,
1178            start: 0,
1179            end: 0,
1180        });
1181        lens.push(clen);
1182        prev_run = first_run;
1183    }
1184    let mut start = pos as u64;
1185    for (e, len) in entries.iter_mut().zip(lens) {
1186        let end = start
1187            .checked_add(len)
1188            .filter(|&e| e <= total_len)
1189            .ok_or(FileError::Container("dict chunk overruns section"))?;
1190        e.start = start;
1191        e.end = end;
1192        start = end;
1193    }
1194    Ok((meta, entries))
1195}
1196
1197/// The `i`-th chunk's first run from the previous chunk's and the stored delta.
1198/// Both writers start chunk 0 at run 0 and give every later chunk at least one
1199/// run, so the runs strictly ascend — which is what routing a term id to its
1200/// chunk (a binary search over `first_run`) relies on. A directory that breaks
1201/// that, or whose sum overflows, is corrupt: refuse it instead of routing ids
1202/// to the wrong chunk.
1203fn next_chunk_run(i: usize, prev_run: usize, drun: usize) -> Result<usize, FileError> {
1204    match (i, drun) {
1205        (0, 0) => Ok(0),
1206        (0, _) => Err(FileError::Container(
1207            "first dict chunk does not start at run 0",
1208        )),
1209        (_, 0) => Err(FileError::Container("dict chunk runs do not ascend")),
1210        _ => prev_run
1211            .checked_add(drun)
1212            .ok_or(FileError::Container("dict chunk run overflows")),
1213    }
1214}
1215
1216/// Fetch and parse a remote chunked dict section's header + directory: read a
1217/// small prefix and grow it geometrically until it parses, never fetching past
1218/// the section.
1219/// Read a chunked dictionary section's directory over a range reader WITHOUT
1220/// materializing the section-wide restart table. That table is one offset per
1221/// restart run — a 50 M-term section has ~3 M of them (~24 MiB resident), and
1222/// holding it is an iOS-Safari OOM on a big remote file. We read only the tiny
1223/// header prefix (term_count / interval) and the chunk directory, skipping the
1224/// restart-table bytes entirely; per-run offsets are derived per chunk on fault
1225/// (`SectionChunk::run_offsets`). The returned meta has an empty
1226/// `restart_offsets`, which the chunked lookups read as "derive per chunk".
1227pub(crate) fn read_dict_dir_ranged<R: RangeReader + ?Sized>(
1228    reader: &R,
1229    section: ByteRange,
1230) -> Result<(crate::dict::SectionMeta, Vec<DictChunkEntry>), FileError> {
1231    let total = section.len;
1232    // Initial prefix: the header prefix ([header_len][term_count][interval]) and,
1233    // for a *small* section, the whole chunk directory too — so those still cost
1234    // a single read. A big section has a huge restart table between the header
1235    // and the directory; we detect that (dir_start past the prefix) and range-
1236    // read only the directory below, never fetching the table.
1237    let init = 8192.min(total); // never over-read past the section (a tiny/empty
1238                                // section holds only its header + a stub directory)
1239    let head = reader.read_at(section.offset, init)?;
1240    let (header_len, n0) =
1241        read_uvarint(&head).ok_or(FileError::Container("truncated dict header len"))?;
1242    let hbase = n0; // first byte of the header body
1243    let (term_count, n1) = read_uvarint(head.get(hbase..).unwrap_or(&[]))
1244        .ok_or(FileError::Container("truncated dict term_count"))?;
1245    let (restart_interval, _n2) = read_uvarint(head.get(hbase + n1..).unwrap_or(&[]))
1246        .ok_or(FileError::Container("truncated dict interval"))?;
1247    if restart_interval == 0 {
1248        return Err(FileError::Container("zero restart interval"));
1249    }
1250    // The chunk directory begins right after the header body — i.e. past the
1251    // `header_len` bytes, which include the restart table we never materialize.
1252    let dir_start = (hbase as u64)
1253        .checked_add(header_len)
1254        .filter(|&d| d <= total)
1255        .ok_or(FileError::Container("dict header overruns section"))?;
1256    let dir_total = total - dir_start;
1257    let meta = crate::dict::SectionMeta {
1258        term_count: u32::try_from(term_count)
1259            .map_err(|_| FileError::Container("dict term count too large"))?,
1260        restart_interval: u32::try_from(restart_interval)
1261            .map_err(|_| FileError::Container("dict restart interval too large"))?,
1262        restart_offsets: Vec::new(),
1263    };
1264    // Runs in the section: what a chunk's first run must stay below.
1265    let num_runs = (meta.term_count as usize).div_ceil(meta.restart_interval as usize);
1266    let finish = |mut entries: Vec<DictChunkEntry>| {
1267        for e in &mut entries {
1268            e.start += dir_start; // dir-relative → section-relative
1269            e.end += dir_start;
1270        }
1271        (meta.clone(), entries)
1272    };
1273    // Fast path: the directory already sits in the prefix we read (small section
1274    // — its restart table is tiny, so the ~few KiB over-read is negligible).
1275    // A short prefix is not wasted: those bytes seed the probe below.
1276    let mut have: Vec<u8> = Vec::new();
1277    if dir_start < head.len() as u64 {
1278        match parse_chunk_dir_only(&head[dir_start as usize..], dir_total, num_runs)? {
1279            ChunkDirParse::Done(entries) => return Ok(finish(entries)),
1280            ChunkDirParse::Truncated { .. } => have = head[dir_start as usize..].to_vec(),
1281        }
1282    }
1283    // Big section: range-read the directory on its own, skipping the table.
1284    //
1285    // Nothing records the directory's byte length, so it has to be probed. Two
1286    // things keep that probe near ONE directory's worth of bytes. Each round
1287    // **appends** to the bytes already held instead of re-reading from the
1288    // start — the previous loop re-read the whole prefix every time, so a
1289    // 234 MB directory (epfl-infoscience's object section) cost 537 MB of range
1290    // reads to fetch. And a truncated parse reports how many entries fit in how
1291    // many bytes, which extrapolates the rest instead of blindly doubling. The
1292    // extrapolation may only quadruple a round, so a wild guess on a directory
1293    // with wildly uneven entries (one stored term can be hundreds of KB) can
1294    // never fetch the section body.
1295    let mut want = 4096u64.min(dir_total).max(1);
1296    loop {
1297        let held = have.len() as u64;
1298        if want > held {
1299            let extra = reader.read_at(section.offset + dir_start + held, want - held)?;
1300            if extra.is_empty() {
1301                return Err(FileError::Container("truncated dict chunk directory"));
1302            }
1303            have.extend_from_slice(&extra);
1304        }
1305        let held = have.len() as u64;
1306        match parse_chunk_dir_only(&have, dir_total, num_runs)? {
1307            ChunkDirParse::Done(entries) => return Ok(finish(entries)),
1308            ChunkDirParse::Truncated { .. } if held >= dir_total => {
1309                return Err(FileError::Container("truncated dict chunk directory"));
1310            }
1311            ChunkDirParse::Truncated {
1312                parsed,
1313                used,
1314                total,
1315            } => {
1316                // Extrapolate from a sample big enough to mean something;
1317                // otherwise just double. Undershooting only costs another
1318                // (append-only) round, so the estimate carries a small margin
1319                // rather than a generous one.
1320                let est = if parsed >= 16 && used > 0 {
1321                    let whole = ((used as u64) / (parsed as u64)).saturating_mul(total as u64);
1322                    whole.saturating_add(whole / 8).saturating_add(64)
1323                } else {
1324                    // Too small a sample to extrapolate from: double.
1325                    held.saturating_mul(2)
1326                };
1327                want = est
1328                    .max(held.saturating_add(1))
1329                    .min(held.saturating_mul(4))
1330                    .min(dir_total);
1331            }
1332        }
1333    }
1334}
1335
1336/// What [`parse_chunk_dir_only`] made of a chunk-directory prefix: either the
1337/// whole directory, or how far a truncated one got (the probe's step hint).
1338/// Truncation is not an error here — the caller is deliberately reading a
1339/// prefix and deciding how much more to fetch.
1340enum ChunkDirParse {
1341    Done(Vec<DictChunkEntry>),
1342    Truncated {
1343        /// Entries fully decoded from the prefix.
1344        parsed: usize,
1345        /// Bytes those entries occupied (directory-relative).
1346        used: usize,
1347        /// Chunk count the directory declares.
1348        total: usize,
1349    },
1350}
1351
1352/// Parse just the chunk directory (the bytes after a section header):
1353/// `[num_chunks][per chunk: Δfirst_run, key_len, key, comp_len]`.
1354/// Chunk byte ranges (`start`/`end`) come back relative to the directory's own
1355/// start; `body_start` is 0 (a lite section never uses it — lookups derive run
1356/// offsets per chunk). Bodies aren't needed here, so `dir` may end at the first
1357/// body as long as it covers the whole directory. A `dir` that stops mid-entry
1358/// is reported as [`ChunkDirParse::Truncated`], not an error: only bytes that
1359/// cannot be a directory at any length (a chunk range overrunning the section,
1360/// chunk runs that do not ascend from 0 or pass `num_runs`) fail.
1361fn parse_chunk_dir_only(
1362    dir: &[u8],
1363    dir_total: u64,
1364    num_runs: usize,
1365) -> Result<ChunkDirParse, FileError> {
1366    let mut pos = 0usize;
1367    let take = |pos: &mut usize| -> Option<u64> {
1368        let (v, n) = read_uvarint(dir.get(*pos..).unwrap_or(&[]))?;
1369        *pos += n;
1370        Some(v)
1371    };
1372    let Some(num_chunks) = take(&mut pos) else {
1373        return Ok(ChunkDirParse::Truncated {
1374            parsed: 0,
1375            used: 0,
1376            total: 0,
1377        });
1378    };
1379    let num_chunks = wire_usize(num_chunks, "dict chunk count too large")?;
1380    let mut entries = Vec::with_capacity(num_chunks.min(dir.len()));
1381    let mut lens: Vec<u64> = Vec::with_capacity(num_chunks.min(dir.len()));
1382    let mut prev_run = 0usize;
1383    for i in 0..num_chunks {
1384        let entry_start = pos;
1385        let short = ChunkDirParse::Truncated {
1386            parsed: entries.len(),
1387            used: entry_start,
1388            total: num_chunks,
1389        };
1390        let (Some(drun), Some(klen)) = (take(&mut pos), take(&mut pos)) else {
1391            return Ok(short);
1392        };
1393        let drun = wire_usize(drun, "dict chunk run too large")?;
1394        let klen = wire_usize(klen, "dict chunk key too large")?;
1395        let Some(key) = dir.get(pos..pos.saturating_add(klen)) else {
1396            return Ok(short);
1397        };
1398        let key = key.to_vec();
1399        pos += klen;
1400        let Some(clen) = take(&mut pos) else {
1401            return Ok(short);
1402        };
1403        let first_run = next_chunk_run(i, prev_run, drun)?;
1404        if first_run >= num_runs {
1405            return Err(FileError::Container("dict chunk run out of range"));
1406        }
1407        entries.push(DictChunkEntry {
1408            first_run,
1409            key,
1410            body_start: 0,
1411            start: 0,
1412            end: 0,
1413        });
1414        lens.push(clen);
1415        prev_run = first_run;
1416    }
1417    let mut start = pos as u64;
1418    for (e, len) in entries.iter_mut().zip(lens) {
1419        let end = start
1420            .checked_add(len)
1421            .filter(|&e| e <= dir_total)
1422            .ok_or(FileError::Container("dict chunk overruns section"))?;
1423        e.start = start;
1424        e.end = end;
1425        start = end;
1426    }
1427    Ok(ChunkDirParse::Done(entries))
1428}
1429
1430/// Decode one chunked dictionary section payload into a resident
1431/// [`crate::dict::ChunkedSection`] (chunks decompressed up front — the local
1432/// open path).
1433fn decode_chunked_dict_section(
1434    payload: &[u8],
1435    codec: u8,
1436    cache: std::sync::Arc<crate::chunk_cache::ChunkCache>,
1437    section_index: u8,
1438) -> Result<crate::dict::ChunkedSection, FileError> {
1439    let (meta, entries) = parse_chunked_dict_dir(payload, payload.len() as u64)?;
1440    let chunks = entries
1441        .into_iter()
1442        .map(|e| {
1443            let body = decompress(codec, &payload[e.start as usize..e.end as usize])?;
1444            Ok((
1445                crate::dict::SectionChunk::new(e.first_run, e.key, e.body_start),
1446                body,
1447            ))
1448        })
1449        .collect::<Result<Vec<_>, FileError>>()?;
1450    Ok(crate::dict::ChunkedSection::resident(
1451        meta,
1452        chunks,
1453        cache,
1454        section_index,
1455    ))
1456}
1457
1458fn decode_dictionary_container(bytes: &[u8], codec: u8) -> Result<Dictionary, FileError> {
1459    let dsecs = decode_container(bytes, CODEC_NONE)?;
1460    if dsecs.len() != 4 {
1461        return Err(FileError::Container("expected 4 dictionary sections"));
1462    }
1463    // The four sections share one (unlimited) cache — residency is identical to
1464    // the previous per-chunk `OnceLock`s.
1465    let cache = crate::chunk_cache::ChunkCache::unlimited_arc();
1466    let mut sections = Vec::with_capacity(4);
1467    for (si, sec) in dsecs.iter().enumerate() {
1468        sections.push(decode_chunked_dict_section(
1469            sec,
1470            codec,
1471            cache.clone(),
1472            si as u8,
1473        )?);
1474    }
1475    let arr: [crate::dict::ChunkedSection; 4] = sections
1476        .try_into()
1477        .map_err(|_| FileError::Container("expected 4 dictionary sections"))?;
1478    Ok(Dictionary::from_chunked_sections(arr))
1479}
1480
1481/// Encode one permutation's tiled section payload (format v0.2):
1482/// `[num_tiles][per tile: delta(min_a), max_a - min_a, compressed_len][tiles…]`,
1483/// each tile compressed independently with `codec` so a ranged reader can
1484/// fetch and decompress exactly the tiles a query routes to. The directory
1485/// itself is uncompressed (it must be readable before any tile).
1486fn encode_tiled_section(index: &GraphIndex, si: usize, codec: u8) -> Vec<u8> {
1487    let tiles = index.tile_sections()[si];
1488    let n = tiles.len();
1489    // Tile bodies live in the index's chunk cache (keyed by (section, tile));
1490    // fetch each as a shared handle. For a locally-built index every tile is
1491    // resident, so this is a cheap `Arc` clone under an uncontended lock.
1492    // Per-tile compression is the bulk of serialization time on a large graph and
1493    // the tiles are independent, so compress them across all cores. `par_iter`
1494    // preserves order, so the output is byte-identical to the serial map.
1495    #[cfg(feature = "parallel")]
1496    let compressed: Vec<Vec<u8>> = {
1497        use rayon::prelude::*;
1498        (0..n)
1499            .into_par_iter()
1500            .map(|ti| compress(codec, &index.tile_body(si, ti)))
1501            .collect()
1502    };
1503    #[cfg(not(feature = "parallel"))]
1504    let compressed: Vec<Vec<u8>> = (0..n)
1505        .map(|ti| compress(codec, &index.tile_body(si, ti)))
1506        .collect();
1507    let mut out = Vec::new();
1508    write_uvarint(&mut out, tiles.len() as u64);
1509    let mut prev_min = 0u32;
1510    for (tile, comp) in tiles.iter().zip(&compressed) {
1511        let (min_a, max_a) = tile.leading_range();
1512        write_uvarint(&mut out, (min_a - prev_min) as u64);
1513        write_uvarint(&mut out, (max_a - min_a) as u64);
1514        write_uvarint(&mut out, comp.len() as u64);
1515        prev_min = min_a;
1516    }
1517    for comp in &compressed {
1518        out.extend_from_slice(comp);
1519    }
1520    // Tile-synopsis trailer (FLAG_TILE_SYNOPSIS): per tile, the inclusive min/max
1521    // of the two non-leading columns `(min_b, span_b, min_c, span_c)`, derived
1522    // from each tile's zone map. Appended **after** the tile payloads so a reader
1523    // that predates the flag — which locates tiles by length and stops — never
1524    // reads it (backward-compatible). A reader honoring the flag reads it from the
1525    // section tail. On the (impossible for a built tile) parse failure, emit a
1526    // full range so nothing is ever wrongly pruned.
1527    for ti in 0..n {
1528        let body = index.tile_body(si, ti);
1529        let (min_b, max_b, min_c, max_c) = match crate::triples::TripleBlock::parse(&body) {
1530            Ok(b) => {
1531                let z = b.zone();
1532                (z.min_b, z.max_b, z.min_c, z.max_c)
1533            }
1534            Err(_) => (0, u32::MAX, 0, u32::MAX),
1535        };
1536        write_uvarint(&mut out, min_b as u64);
1537        write_uvarint(&mut out, (max_b - min_b) as u64);
1538        write_uvarint(&mut out, min_c as u64);
1539        write_uvarint(&mut out, (max_c - min_c) as u64);
1540    }
1541    out
1542}
1543
1544/// A parsed v0.2 tile-directory entry: leading-id range plus the tile's byte
1545/// range *within the section payload*.
1546pub(crate) struct TileDirEntry {
1547    pub(crate) min_a: u32,
1548    pub(crate) max_a: u32,
1549    pub(crate) start: u64,
1550    pub(crate) end: u64,
1551}
1552
1553/// One tile's synopsis: inclusive min/max of the two non-leading columns.
1554type TileSynopsis = (u32, u32, u32, u32);
1555
1556/// Parse the **tile-synopsis trailer** (when [`FLAG_TILE_SYNOPSIS`] is set): the
1557/// `num_tiles × (min_b, span_b, min_c, span_c)` records that follow the last tile
1558/// payload, starting at `trailer_start` within `payload`. Returns one synopsis per
1559/// tile, in directory order. `payload` may be just the trailer slice (remote) or
1560/// the whole section (local); `trailer_start` is the offset of the trailer within
1561/// it. A short/garbled trailer yields `None` (the caller keeps `None` synopses —
1562/// pruning simply doesn't fire, never a wrong result).
1563fn parse_tile_synopsis(
1564    payload: &[u8],
1565    trailer_start: usize,
1566    num_tiles: usize,
1567) -> Option<Vec<TileSynopsis>> {
1568    let mut pos = trailer_start;
1569    let take = |pos: &mut usize| -> Option<u32> {
1570        let (v, n) = read_uvarint(payload.get(*pos..)?)?;
1571        *pos += n;
1572        u32::try_from(v).ok()
1573    };
1574    let mut out = Vec::with_capacity(num_tiles.min(payload.len()));
1575    for _ in 0..num_tiles {
1576        let min_b = take(&mut pos)?;
1577        let max_b = min_b.checked_add(take(&mut pos)?)?;
1578        let min_c = take(&mut pos)?;
1579        let max_c = min_c.checked_add(take(&mut pos)?)?;
1580        out.push((min_b, max_b, min_c, max_c));
1581    }
1582    Some(out)
1583}
1584
1585/// Parse a tiled section payload's directory (not the tiles). `bytes` may be a
1586/// **prefix** of the payload (a ranged reader fetches the directory before any
1587/// tile); tile byte ranges are validated against `total_len`, the full payload
1588/// length. Every length is untrusted.
1589fn parse_tile_directory(bytes: &[u8], total_len: u64) -> Result<Vec<TileDirEntry>, FileError> {
1590    let mut pos = 0usize;
1591    let take = |pos: &mut usize| -> Result<u64, FileError> {
1592        let (v, n) = read_uvarint(bytes.get(*pos..).unwrap_or(&[]))
1593            .ok_or(FileError::Container("truncated tile directory"))?;
1594        *pos += n;
1595        Ok(v)
1596    };
1597    let num_tiles = wire_usize(take(&mut pos)?, "tile count too large")?;
1598    let mut entries = Vec::with_capacity(num_tiles.min(bytes.len()));
1599    let mut prev_min = 0u32;
1600    let mut lens = Vec::with_capacity(num_tiles.min(bytes.len()));
1601    // Leading ids are u32. A delta or span past that, or a sum that leaves
1602    // the id space, is corrupt: `as u32` and `wrapping_add` used to turn it
1603    // into some other, plausible id range and route lookups to the wrong tile.
1604    let id = |v: u64| u32::try_from(v).map_err(|_| FileError::Container("tile id out of range"));
1605    for _ in 0..num_tiles {
1606        let dmin = id(take(&mut pos)?)?;
1607        let span = id(take(&mut pos)?)?;
1608        let len = take(&mut pos)?;
1609        let min_a = prev_min
1610            .checked_add(dmin)
1611            .ok_or(FileError::Container("tile id out of range"))?;
1612        let max_a = min_a
1613            .checked_add(span)
1614            .ok_or(FileError::Container("tile id out of range"))?;
1615        entries.push(TileDirEntry {
1616            min_a,
1617            max_a,
1618            start: 0,
1619            end: 0,
1620        });
1621        lens.push(len);
1622        prev_min = min_a;
1623    }
1624    let mut start = pos as u64;
1625    for (e, len) in entries.iter_mut().zip(lens) {
1626        let end = start
1627            .checked_add(len)
1628            .filter(|&e| e <= total_len)
1629            .ok_or(FileError::Container("tile overruns section"))?;
1630        e.start = start;
1631        e.end = end;
1632        start = end;
1633    }
1634    Ok(entries)
1635}
1636
1637/// Fetch and parse a remote tiled section's directory: read a small prefix and
1638/// grow it geometrically until the directory parses, never fetching past the
1639/// section. A directory that still fails on the whole section is corrupt.
1640pub(crate) fn read_tile_directory_ranged<R: RangeReader + ?Sized>(
1641    reader: &R,
1642    section: ByteRange,
1643) -> Result<Vec<TileDirEntry>, FileError> {
1644    let total = section.len;
1645    let mut prefetch = 4096u64.min(total);
1646    loop {
1647        let prefix = reader.read_at(section.offset, prefetch)?;
1648        match parse_tile_directory(&prefix, total) {
1649            Ok(dir) => return Ok(dir),
1650            Err(_) if prefetch < total => prefetch = prefetch.saturating_mul(2).min(total),
1651            Err(e) => return Err(e),
1652        }
1653    }
1654}
1655
1656/// Fetch and parse a remote section's **tile-synopsis trailer** (only when the
1657/// header's [`FLAG_TILE_SYNOPSIS`] is set): one targeted range read of the bytes
1658/// past the last tile, parsed into one synopsis per tile (directory order). A
1659/// missing/short/garbled trailer degrades to all-`None` — pruning simply doesn't
1660/// fire, never a wrong result. The directory gives the trailer's start (the last
1661/// tile's end).
1662pub(crate) fn read_tile_synopsis_ranged<R: RangeReader + ?Sized>(
1663    reader: &R,
1664    section: ByteRange,
1665    dir: &[TileDirEntry],
1666) -> Vec<Option<TileSynopsis>> {
1667    let n = dir.len();
1668    let none = vec![None; n];
1669    let trailer_start = dir.iter().map(|e| e.end).max().unwrap_or(0);
1670    let total = section.len;
1671    if n == 0 || trailer_start >= total {
1672        return none; // no trailer bytes present
1673    }
1674    // Four varints per tile, each a u32 (at most 5 bytes): never read more,
1675    // whatever the section claims. The trailer is the section's tail, so a
1676    // hostile section length used to make this one read as large as it said.
1677    let max_trailer = (n as u64).saturating_mul(4 * 5);
1678    let trailer_len = (total - trailer_start).min(max_trailer);
1679    let Some(offset) = section.offset.checked_add(trailer_start) else {
1680        return none;
1681    };
1682    let Ok(bytes) = reader.read_at(offset, trailer_len) else {
1683        return none;
1684    };
1685    match parse_tile_synopsis(&bytes, 0, n) {
1686        Some(v) => v.into_iter().map(Some).collect(),
1687        None => none,
1688    }
1689}
1690
1691/// Per-tile absolute file ranges of each permutation section, for provenance.
1692/// A malformed directory yields an empty section (provenance degrades, queries
1693/// are unaffected).
1694fn tile_file_ranges(
1695    index_bytes: &[u8],
1696    container_offset: u64,
1697    section_ranges: &[ByteRange; NUM_PERMS],
1698) -> [Vec<(u32, u32, ByteRange)>; NUM_PERMS] {
1699    let mut out: [Vec<(u32, u32, ByteRange)>; NUM_PERMS] = Default::default();
1700    for (section, range) in out.iter_mut().zip(section_ranges) {
1701        // A permutation the file does not carry has a zeroed range. Its
1702        // `offset - container_offset` underflows, and before this guard it
1703        // panicked the whole `open_ranged` path on the first lean file.
1704        if range.len == 0 || range.offset < container_offset {
1705            continue;
1706        }
1707        let Some(payload) = checked_slice(index_bytes, range.offset - container_offset, range.len)
1708        else {
1709            continue;
1710        };
1711        if let Ok(dir) = parse_tile_directory(payload, payload.len() as u64) {
1712            *section = dir
1713                .into_iter()
1714                .map(|e| {
1715                    (
1716                        e.min_a,
1717                        e.max_a,
1718                        ByteRange {
1719                            offset: range.offset + e.start,
1720                            len: (e.end - e.start),
1721                        },
1722                    )
1723                })
1724                .collect();
1725        }
1726    }
1727    out
1728}
1729
1730/// Decode a tiled section payload into `(min_a, max_a, uncompressed tile)`
1731/// triples.
1732fn decode_tiled_section(payload: &[u8], codec: u8) -> Result<Vec<(u32, u32, Vec<u8>)>, FileError> {
1733    parse_tile_directory(payload, payload.len() as u64)?
1734        .into_iter()
1735        .map(|e| {
1736            Ok((
1737                e.min_a,
1738                e.max_a,
1739                decompress(codec, &payload[e.start as usize..e.end as usize])?,
1740            ))
1741        })
1742        .collect()
1743}
1744
1745/// Decode the index container: one raw tiled section payload per permutation
1746/// the file carries (`perms`, from the header — [`PermSet::ALL`] for every file
1747/// written before the mask existed), each tile compressed individually.
1748///
1749/// The section count is checked against `perms.len()`, so a file whose header
1750/// and container disagree is rejected rather than silently short-read. A reader
1751/// that does not know about lean files at all reaches the same conclusion by a
1752/// different route: it passes six, the container says three, and this errors —
1753/// which is why a lean file cannot be misread as an empty one.
1754fn decode_index_container(
1755    bytes: &[u8],
1756    codec: u8,
1757    perms: PermSet,
1758) -> Result<GraphIndex, FileError> {
1759    let isecs = decode_container(bytes, CODEC_NONE)?;
1760    if isecs.len() != perms.len() {
1761        return Err(FileError::Container(
1762            "index container section count does not match the header permutation mask",
1763        ));
1764    }
1765    decode_index_sections(isecs, codec, perms)
1766}
1767
1768fn decode_index_sections(
1769    mut isecs: Vec<Vec<u8>>,
1770    codec: u8,
1771    perms: PermSet,
1772) -> Result<GraphIndex, FileError> {
1773    let mut sections: [Vec<(u32, u32, Vec<u8>)>; NUM_PERMS] = Default::default();
1774    for (perm, sec) in perms.iter().zip(isecs.iter_mut()) {
1775        sections[perm.section_index()] = decode_tiled_section(sec, codec)?;
1776    }
1777    Ok(GraphIndex::from_tiles(sections, perms))
1778}
1779
1780/// Decode the file's **root** index container (the default graph): the
1781/// permutation sections, plus — on a quad-index file — the trailing seal, which
1782/// is checked and set aside. Returns the seal's digest of the QUAD_INDEX
1783/// section when there is one.
1784fn decode_root_index_container(
1785    bytes: &[u8],
1786    header: &Header,
1787) -> Result<(GraphIndex, Option<[u8; 16]>), FileError> {
1788    if !header.has_quad_index() {
1789        return Ok((
1790            decode_index_container(bytes, header.block_codec, header.perms)?,
1791            None,
1792        ));
1793    }
1794    let mut isecs = decode_container(bytes, CODEC_NONE)?;
1795    if isecs.len() != header.index_container_sections() {
1796        return Err(FileError::Container(
1797            "index container section count does not match the header permutation mask",
1798        ));
1799    }
1800    let seal = isecs.pop().unwrap_or_default();
1801    let digest = crate::quad::parse_seal(&seal)?;
1802    Ok((
1803        decode_index_sections(isecs, header.block_codec, header.perms)?,
1804        Some(digest),
1805    ))
1806}
1807
1808/// The named-graph layout a header announces must be the one its directory
1809/// holds: the quad flag with a QUAD_INDEX section and no NAMED_GRAPHS section,
1810/// or no flag and no QUAD_INDEX section. Anything else is a corrupt (or a
1811/// hand-edited) header, refused before either layout is read.
1812fn check_named_layout(header: &Header) -> Result<(), FileError> {
1813    let quad = header.quad_index_section().is_some();
1814    if header.has_quad_index() {
1815        if !quad {
1816            return Err(FileError::Container(
1817                "quad-index flag set but the file has no quad-index section",
1818            ));
1819        }
1820        if header.named_graphs_len > 0 {
1821            return Err(FileError::Container(
1822                "a file cannot carry both a quad index and a named-graphs section",
1823            ));
1824        }
1825        crate::quad::check_perms(header.perms)?;
1826    } else if quad {
1827        return Err(FileError::Container(
1828            "quad-index section present but the header does not flag it",
1829        ));
1830    }
1831    Ok(())
1832}
1833
1834fn container_section_payload_ranges(
1835    bytes: &[u8],
1836    container_offset: u64,
1837    expected_sections: usize,
1838) -> Result<Vec<ByteRange>, FileError> {
1839    let (section_count, mut pos) =
1840        read_uvarint(bytes).ok_or(FileError::Container("truncated count"))?;
1841    let section_count = usize::try_from(section_count)
1842        .map_err(|_| FileError::Container("section count too large"))?;
1843    if section_count != expected_sections {
1844        return Err(FileError::Container("unexpected section count"));
1845    }
1846
1847    let mut ranges = Vec::with_capacity(section_count);
1848    for _ in 0..section_count {
1849        let remaining = bytes
1850            .get(pos..)
1851            .ok_or(FileError::Container("truncated length"))?;
1852        let (payload_len, used) =
1853            read_uvarint(remaining).ok_or(FileError::Container("truncated length"))?;
1854        pos = pos
1855            .checked_add(used)
1856            .ok_or(FileError::Container("section range overflows"))?;
1857        let payload_len_usize = usize::try_from(payload_len)
1858            .map_err(|_| FileError::Container("section length too large"))?;
1859        let payload_end = pos
1860            .checked_add(payload_len_usize)
1861            .ok_or(FileError::Container("section range overflows"))?;
1862        if payload_end > bytes.len() {
1863            return Err(FileError::Container("section overruns buffer"));
1864        }
1865        ranges.push(ByteRange {
1866            offset: checked_end(container_offset, pos as u64)?,
1867            len: payload_len,
1868        });
1869        pos = payload_end;
1870    }
1871
1872    Ok(ranges)
1873}
1874
1875/// The permutation payload ranges of an index container holding `sections`
1876/// sections (the permutations, then — on a quad-index root — the seal, which
1877/// is not a permutation and is left out).
1878fn decode_index_section_ranges(
1879    bytes: &[u8],
1880    container_offset: u64,
1881    perms: PermSet,
1882    sections: usize,
1883) -> Result<[ByteRange; NUM_PERMS], FileError> {
1884    let ranges = container_section_payload_ranges(bytes, container_offset, sections)?;
1885    let mut out = [ByteRange { offset: 0, len: 0 }; NUM_PERMS];
1886    for (perm, range) in perms.iter().zip(ranges) {
1887        out[perm.section_index()] = range;
1888    }
1889    Ok(out)
1890}
1891
1892fn read_uvarint_at<R: RangeReader + ?Sized>(
1893    reader: &R,
1894    absolute_offset: u64,
1895    container_end: u64,
1896) -> Result<(u64, u64), FileError> {
1897    if absolute_offset >= container_end {
1898        return Err(FileError::Container("truncated container varint"));
1899    }
1900    let remaining = container_end - absolute_offset;
1901    let probe_len = remaining.min(10);
1902    let bytes = reader.read_at(absolute_offset, probe_len)?;
1903    read_uvarint(&bytes)
1904        .map(|(value, used)| (value, used as u64))
1905        .ok_or(FileError::Container("truncated container varint"))
1906}
1907
1908/// Locate one section's payload byte range inside a remote container, walking
1909/// only the (tiny) varint framing — no payload bytes are fetched.
1910pub(crate) fn locate_container_section_ranged<R: RangeReader + ?Sized>(
1911    reader: &R,
1912    container_offset: u64,
1913    container_len: u64,
1914    section_index: usize,
1915    expected_sections: u64,
1916) -> Result<ByteRange, FileError> {
1917    let container_end = checked_end(container_offset, container_len)?;
1918    let (section_count, used) = read_uvarint_at(reader, container_offset, container_end)?;
1919    if section_count != expected_sections {
1920        return Err(FileError::Container("unexpected container section count"));
1921    }
1922    if section_index >= section_count as usize {
1923        return Err(FileError::Container(
1924            "container section index out of bounds",
1925        ));
1926    }
1927
1928    let mut pos = checked_end(container_offset, used)?;
1929    for i in 0..section_count as usize {
1930        let (payload_len, len_used) = read_uvarint_at(reader, pos, container_end)?;
1931        pos = checked_end(pos, len_used)?;
1932        let payload_end = checked_end(pos, payload_len)?;
1933        if payload_end > container_end {
1934            return Err(FileError::Container("section overruns buffer"));
1935        }
1936        if i == section_index {
1937            return Ok(ByteRange {
1938                offset: pos,
1939                len: payload_len,
1940            });
1941        }
1942        pos = payload_end;
1943    }
1944    Err(FileError::Container("container section not found"))
1945}
1946
1947/// Open one index container (six tiled permutation sections) **lazily** over a
1948/// range reader: fetch each section's tile directory (and synopsis trailer,
1949/// when the file has one) but no tile payloads — tiles fault in on first scan.
1950/// Returns the index plus its section and tile file ranges (provenance).
1951///
1952/// This is the machinery `open_ranged_lazy` always used for the default
1953/// graph's container at `root_dir_offset`; a named graph's container is the
1954/// same format at a different offset, so the lazy named-graphs path opens a
1955/// LARGE graph through this too instead of decoding it resident.
1956#[allow(clippy::type_complexity)]
1957fn open_index_container_lazy(
1958    reader: &std::sync::Arc<dyn RangeReader + Send + Sync>,
1959    container: ByteRange,
1960    block_codec: u8,
1961    has_synopsis: bool,
1962    read_concurrency: usize,
1963    perms: PermSet,
1964    sections: usize,
1965) -> Result<
1966    (
1967        GraphIndex,
1968        [ByteRange; NUM_PERMS],
1969        [Vec<(u32, u32, ByteRange)>; NUM_PERMS],
1970    ),
1971    FileError,
1972> {
1973    let mut index_section_ranges = [ByteRange { offset: 0, len: 0 }; NUM_PERMS];
1974    let mut tile_ranges: [Vec<(u32, u32, ByteRange)>; NUM_PERMS] = Default::default();
1975    #[allow(clippy::type_complexity)]
1976    let mut directories: [Vec<(u32, u32, Option<TileSynopsis>)>; NUM_PERMS] = Default::default();
1977    for (pos, perm) in perms.iter().enumerate() {
1978        let si = perm.section_index();
1979        let section = locate_container_section_ranged(
1980            reader,
1981            container.offset,
1982            container.len,
1983            pos,
1984            sections as u64,
1985        )?;
1986        index_section_ranges[si] = section;
1987        let dir = read_tile_directory_ranged(reader, section)?;
1988        // Tile synopses (one extra small tail read per section) let a routed
1989        // scan prune a tile by a bound secondary component before faulting it.
1990        let syn = if has_synopsis {
1991            read_tile_synopsis_ranged(reader, section, &dir)
1992        } else {
1993            vec![None; dir.len()]
1994        };
1995        directories[si] = dir
1996            .iter()
1997            .zip(syn)
1998            .map(|(e, s)| (e.min_a, e.max_a, s))
1999            .collect();
2000        tile_ranges[si] = dir
2001            .into_iter()
2002            .map(|e| {
2003                (
2004                    e.min_a,
2005                    e.max_a,
2006                    ByteRange {
2007                        offset: section.offset + e.start,
2008                        len: (e.end - e.start),
2009                    },
2010                )
2011            })
2012            .collect();
2013    }
2014
2015    // The loader fetches and decompresses one tile per call; the bulk
2016    // loader serves multi-tile scans by coalescing adjacent tile ranges
2017    // into single range reads (tiles are back-to-back in their section,
2018    // so a full-section scan is typically one request).
2019    //
2020    // Either loader returns `None` (a failed load: the index marks itself
2021    // incomplete and a reset retries) unless the tile arrived whole and its
2022    // zone map parses. Without the length check, a short read of a raw
2023    // (uncompressed) tile decoded as a shorter tile and silently lost rows.
2024    let codec = block_codec;
2025    let loader_ranges = tile_ranges.clone();
2026    let loader_reader = reader.clone();
2027    let loader: crate::index::TileLoader = Box::new(move |si, ti| {
2028        let (_, _, range) = loader_ranges.get(si)?.get(ti)?;
2029        let bytes = loader_reader.read_at(range.offset, range.len).ok()?;
2030        if bytes.len() as u64 != range.len {
2031            return None;
2032        }
2033        checked_tile(codec, &bytes)
2034    });
2035    let bulk_ranges = tile_ranges.clone();
2036    let bulk_reader = reader.clone();
2037    let bulk: crate::index::TileBulkLoader = Box::new(move |si, tis| {
2038        let section = bulk_ranges.get(si)?;
2039        let want: Option<Vec<ByteRange>> = tis
2040            .iter()
2041            .map(|&ti| section.get(ti).map(|&(_, _, r)| r))
2042            .collect();
2043        let blobs = read_coalesced(bulk_reader.as_ref(), &want?, TILE_COALESCE_GAP)?;
2044        blobs.iter().map(|b| checked_tile(codec, b)).collect()
2045    });
2046    let mut index =
2047        GraphIndex::from_remote_directories(directories, perms, loader).with_bulk_loader(bulk);
2048    // Per-tile encoded lengths (from the directory) feed the join planner's
2049    // fatness gates — free here, unavailable later without a fetch.
2050    index.set_tile_lens(std::array::from_fn(|si| {
2051        tile_ranges[si]
2052            .iter()
2053            .map(|&(_, _, r)| r.len.min(u32::MAX as u64) as u32)
2054            .collect()
2055    }));
2056    // The reader's fan-out widens the planner's remote probe budget: a
2057    // desktop/CLI reader overlapping 16 range reads probes far more cheaply
2058    // than a phone's serial sync-XHR path.
2059    index.set_read_concurrency(read_concurrency);
2060    Ok((index, index_section_ranges, tile_ranges))
2061}
2062
2063/// A fetched tile, decompressed, or `None` when it does not decompress or its
2064/// zone map does not parse: a lazy tile load fails on the first scan that
2065/// touches it, instead of a corrupt tile reading as an empty or partial one.
2066fn checked_tile(codec: u8, bytes: &[u8]) -> Option<Vec<u8>> {
2067    let tile = decompress(codec, bytes).ok()?;
2068    crate::triples::TripleBlock::parse(&tile).ok()?;
2069    Some(tile)
2070}
2071
2072/// Serialize a complete `.rete` file image from a dictionary, index, and an
2073/// (optionally empty) encoded pyramid-meta section. `pyramid_levels` records the
2074/// number of dendrogram rounds the pyramid spans (0 if no pyramid).
2075///
2076/// The image is format `0x06`, the one layout every writer produces (the
2077/// legacy `0x05` image is [`write_file_v05`]).
2078pub fn write_file(
2079    dict: &Dictionary,
2080    index: &GraphIndex,
2081    has_quads: bool,
2082    pyramid_meta: &[u8],
2083    pyramid_levels: u16,
2084) -> Vec<u8> {
2085    write_dataset(dict, index, &[], has_quads, pyramid_meta, pyramid_levels)
2086}
2087
2088/// The legacy `0x05` image of [`write_file`]. The layout of the first stable
2089/// generation, kept for the tests that pin its byte layout, for the in-RAM
2090/// builders that hand their `0x05` image to the one `0x06` finish step, and for
2091/// `RETE_WRITE_FORMAT=5`.
2092#[doc(hidden)]
2093pub fn write_file_v05(
2094    dict: &Dictionary,
2095    index: &GraphIndex,
2096    has_quads: bool,
2097    pyramid_meta: &[u8],
2098    pyramid_levels: u16,
2099) -> Vec<u8> {
2100    write_dataset_v05(dict, index, &[], has_quads, pyramid_meta, pyramid_levels)
2101}
2102
2103/// A finished `0x05` image as the `0x06` file the public writers return (the
2104/// image is untouched under `RETE_WRITE_FORMAT=5`).
2105fn into_v06(image: Vec<u8>) -> Vec<u8> {
2106    crate::build_record::finish_in_ram(image, None)
2107        .unwrap_or_else(|e| panic!("writing format 0x06: {e}"))
2108}
2109
2110/// Encode an index container (v0.2): one raw tiled section payload per
2111/// permutation the index carries, in [`crate::index::ALL_PERMS`] order, tiles
2112/// compressed individually with `codec`. A three-permutation index writes three
2113/// sections — not six with three empty, which would be indistinguishable from
2114/// an empty graph to a reader that does not check the header mask.
2115pub(crate) fn encode_index_container(index: &GraphIndex, codec: u8) -> Vec<u8> {
2116    encode_index_container_sealed(index, codec, None)
2117}
2118
2119/// [`encode_index_container`] with an optional trailing **seal** section —
2120/// the root container of a quad-index file (see `quad.rs`). The seal is what
2121/// makes every reader that predates the quad index refuse the file: they all
2122/// check the container's section count before reading a payload.
2123fn encode_index_container_sealed(index: &GraphIndex, codec: u8, seal: Option<&[u8]>) -> Vec<u8> {
2124    let payloads: Vec<Vec<u8>> = index
2125        .perms()
2126        .iter()
2127        .map(|perm| encode_tiled_section(index, perm.section_index(), codec))
2128        .collect();
2129    let mut refs: Vec<&[u8]> = payloads.iter().map(|p| p.as_slice()).collect();
2130    if let Some(seal) = seal {
2131        refs.push(seal);
2132    }
2133    encode_container(&refs, CODEC_NONE)
2134}
2135
2136/// Encode the named-graphs section: each graph as `(iri, permutation container)`.
2137fn encode_named_graphs(named: &[(String, GraphIndex)], codec: u8) -> Vec<u8> {
2138    let mut out = Vec::new();
2139    write_uvarint(&mut out, named.len() as u64);
2140    for (iri, index) in named {
2141        write_uvarint(&mut out, iri.len() as u64);
2142        out.extend_from_slice(iri.as_bytes());
2143        let container = encode_index_container(index, codec);
2144        write_uvarint(&mut out, container.len() as u64);
2145        out.extend_from_slice(&container);
2146    }
2147    out
2148}
2149
2150fn decode_named_graphs(
2151    bytes: &[u8],
2152    codec: u8,
2153    perms: PermSet,
2154) -> Result<Vec<(String, GraphIndex)>, FileError> {
2155    let (n, mut pos) = read_uvarint(bytes).ok_or(FileError::Container("truncated graph count"))?;
2156    // Bounds-checked slice within this (already bounded) section. Lengths read
2157    // below are untrusted, so every range is validated before indexing.
2158    let bound = |start: usize, len: u64| -> Result<usize, FileError> {
2159        usize::try_from(len)
2160            .ok()
2161            .and_then(|len| start.checked_add(len))
2162            .filter(|&e| e <= bytes.len())
2163            .ok_or(FileError::Container("named-graph field overruns buffer"))
2164    };
2165    let n = wire_usize(n, "named-graph count too large")?;
2166    let mut out = Vec::with_capacity(n.min(bytes.len()));
2167    for _ in 0..n {
2168        let (ilen, u1) = read_uvarint(bytes.get(pos..).unwrap_or(&[]))
2169            .ok_or(FileError::Container("truncated iri len"))?;
2170        pos += u1;
2171        let iend = bound(pos, ilen)?;
2172        let iri = String::from_utf8_lossy(&bytes[pos..iend]).into_owned();
2173        pos = iend;
2174        let (clen, u2) = read_uvarint(bytes.get(pos..).unwrap_or(&[]))
2175            .ok_or(FileError::Container("truncated container len"))?;
2176        pos += u2;
2177        let cend = bound(pos, clen)?;
2178        let index = decode_index_container(&bytes[pos..cend], codec, perms)?;
2179        // Resident, so the count is a pass over tile headers; the planner
2180        // sizes this graph's patterns by it, not by the file's total.
2181        index.set_size_hint(u64::from(index.triple_count()));
2182        out.push((iri, index));
2183        pos = cend;
2184    }
2185    Ok(out)
2186}
2187
2188/// Serialize a full RDF *dataset*: the default-graph index plus zero or more
2189/// named graphs `(iri, index)`, all sharing one dictionary. Format `0x06`; named
2190/// graphs, given per graph here, are stored as the quad index (`0x06` has no
2191/// other layout).
2192pub fn write_dataset(
2193    dict: &Dictionary,
2194    default_index: &GraphIndex,
2195    named: &[(String, GraphIndex)],
2196    has_quads: bool,
2197    pyramid_meta: &[u8],
2198    pyramid_levels: u16,
2199) -> Vec<u8> {
2200    write_dataset_with_metadata(
2201        dict,
2202        default_index,
2203        named,
2204        has_quads,
2205        pyramid_meta,
2206        pyramid_levels,
2207        &[],
2208        &[],
2209    )
2210}
2211
2212/// The legacy `0x05` image of [`write_dataset`] (see [`write_file_v05`]).
2213#[doc(hidden)]
2214pub fn write_dataset_v05(
2215    dict: &Dictionary,
2216    default_index: &GraphIndex,
2217    named: &[(String, GraphIndex)],
2218    has_quads: bool,
2219    pyramid_meta: &[u8],
2220    pyramid_levels: u16,
2221) -> Vec<u8> {
2222    write_dataset_with_metadata_v05(
2223        dict,
2224        default_index,
2225        named,
2226        has_quads,
2227        pyramid_meta,
2228        pyramid_levels,
2229        &[],
2230        &[],
2231    )
2232}
2233
2234/// Serialize a dictionary to its on-file container bytes (4 front-coded, chunked
2235/// sections). Exposed so a low-RAM build can serialize **and drop** the live
2236/// `Dictionary` before building the permutation index — the index build works on
2237/// id-triples and never needs the dictionary.
2238pub(crate) fn encode_dict_container(dict: &Dictionary, codec: u8) -> Vec<u8> {
2239    let raw_sections = dict.sections();
2240    let dict_payloads: Vec<Vec<u8>> = raw_sections
2241        .iter()
2242        .map(|raw| encode_chunked_dict_section(raw, codec))
2243        .collect();
2244    encode_container(
2245        &[
2246            dict_payloads[0].as_slice(),
2247            dict_payloads[1].as_slice(),
2248            dict_payloads[2].as_slice(),
2249            dict_payloads[3].as_slice(),
2250        ],
2251        CODEC_NONE,
2252    )
2253}
2254
2255/// Serialize a dataset with an opaque **metadata** payload occupying the file's
2256/// metadata section (the application layer defines its meaning — the CLI stores a
2257/// JSON Dataset Card there). The section sits immediately after the header and
2258/// before the dictionary, so `metadata_offset` stays at `HEADER_LEN` and every
2259/// downstream section shifts by `metadata.len()`. The payload is folded into the
2260/// `content_hash`, so `verify` covers it and it is tamper-evident.
2261///
2262/// Passing an empty `metadata` is byte-identical to [`write_dataset`]: the section
2263/// is omitted (`metadata_len = 0`, `dictionary_offset = HEADER_LEN`) and the hash
2264/// is computed over exactly the same parts (a zero-length hash update is a no-op).
2265#[allow(clippy::too_many_arguments)]
2266pub fn write_dataset_with_metadata(
2267    dict: &Dictionary,
2268    default_index: &GraphIndex,
2269    named: &[(String, GraphIndex)],
2270    has_quads: bool,
2271    pyramid_meta: &[u8],
2272    pyramid_levels: u16,
2273    metadata: &[u8],
2274    text_index: &[u8],
2275) -> Vec<u8> {
2276    into_v06(write_dataset_with_metadata_v05(
2277        dict,
2278        default_index,
2279        named,
2280        has_quads,
2281        pyramid_meta,
2282        pyramid_levels,
2283        metadata,
2284        text_index,
2285    ))
2286}
2287
2288/// The legacy `0x05` image of [`write_dataset_with_metadata`] (see
2289/// [`write_file_v05`]).
2290#[doc(hidden)]
2291#[allow(clippy::too_many_arguments)]
2292pub fn write_dataset_with_metadata_v05(
2293    dict: &Dictionary,
2294    default_index: &GraphIndex,
2295    named: &[(String, GraphIndex)],
2296    has_quads: bool,
2297    pyramid_meta: &[u8],
2298    pyramid_levels: u16,
2299    metadata: &[u8],
2300    text_index: &[u8],
2301) -> Vec<u8> {
2302    let codec = writer_codec();
2303    let dict_container = encode_dict_container(dict, codec);
2304    // This entry point takes a finished section, so it reads the section's own
2305    // scope block to pick the directory kind (6, or 11 for a partial index) and
2306    // the header's named-graphs bit.
2307    let scope = (!text_index.is_empty())
2308        .then(|| crate::text_index::TextIndex::from_section(text_index, codec).ok())
2309        .flatten()
2310        .map(|ti| ti.scope().clone());
2311    let text_index_partial = scope
2312        .as_ref()
2313        .is_some_and(crate::text_index::TextIndexScope::restricts_predicates);
2314    let text_index_named = scope
2315        .as_ref()
2316        .is_some_and(crate::text_index::TextIndexScope::covers_named_graphs);
2317    write_dataset_from_parts(
2318        &dict_container,
2319        dict.term_count() as u64,
2320        default_index,
2321        NamedPart::PerGraph(named),
2322        has_quads,
2323        dict.has_quoted_triples(),
2324        pyramid_meta,
2325        pyramid_levels,
2326        metadata,
2327        text_index,
2328        text_index_partial,
2329        text_index_named,
2330        codec,
2331    )
2332}
2333
2334/// How the named graphs reach the writer: one index per graph (the
2335/// NAMED_GRAPHS section, kind 5), or an encoded quad index (kind 10).
2336pub(crate) enum NamedPart<'a> {
2337    PerGraph(&'a [(String, GraphIndex)]),
2338    Quad(&'a crate::quad::EncodedQuadIndex),
2339}
2340
2341/// Assemble the final file image from an **already-serialized** dictionary
2342/// container (so the caller can drop the live `Dictionary` before calling this)
2343/// plus the permutation index and optional sections. The byte output is identical
2344/// to serializing the dictionary inline.
2345#[allow(clippy::too_many_arguments)]
2346pub(crate) fn write_dataset_from_parts(
2347    dict_container: &[u8],
2348    term_count: u64,
2349    default_index: &GraphIndex,
2350    named: NamedPart<'_>,
2351    has_quads: bool,
2352    has_quoted_triples: bool,
2353    pyramid_meta: &[u8],
2354    pyramid_levels: u16,
2355    metadata: &[u8],
2356    text_index: &[u8],
2357    text_index_partial: bool,
2358    text_index_named: bool,
2359    codec: u8,
2360) -> Vec<u8> {
2361    // The named graphs: either the per-graph section (kind 5) or the quad index
2362    // (kind 10), whose digest the root container's seal carries. An empty quad
2363    // index is written as no named graphs at all.
2364    let (named_section, quad, named_quads) = match named {
2365        NamedPart::PerGraph(named) => (
2366            if named.is_empty() {
2367                Vec::new()
2368            } else {
2369                encode_named_graphs(named, codec)
2370            },
2371            None,
2372            named
2373                .iter()
2374                .map(|(_, idx)| idx.triple_count() as u64)
2375                .sum::<u64>(),
2376        ),
2377        NamedPart::Quad(q) if q.graph_count == 0 => (Vec::new(), None, 0),
2378        NamedPart::Quad(q) => (Vec::new(), Some(q), q.quad_count),
2379    };
2380    let seal = quad.map(|q| crate::quad::encode_seal(crate::quad::digest(&q.bytes)));
2381    let index_container = encode_index_container_sealed(default_index, codec, seal.as_deref());
2382    // A predicate-restricted index (kind 11) is outside the content hash, like
2383    // kinds 8 and 9, and starts with its own digest instead: a reader older
2384    // than 0.4.0 does not see kind 11, so a hashed one would make its `verify`
2385    // call every such file corrupt. A kind-6 index is hashed, as it always was.
2386    let text_index_partial = text_index_partial && !text_index.is_empty();
2387    let digested;
2388    let text_index: &[u8] = if text_index_partial {
2389        digested = crate::text_index::with_partial_digest(text_index);
2390        &digested
2391    } else {
2392        text_index
2393    };
2394
2395    // The metadata section (if any) sits between the header and the dictionary,
2396    // so the dictionary — and everything after it — shifts forward by its length.
2397    let meta_section_len = metadata.len() as u64;
2398    let dict_offset = HEADER_LEN as u64 + meta_section_len;
2399    let dict_len = dict_container.len() as u64;
2400    let index_offset = dict_offset + dict_len;
2401    let index_len = index_container.len() as u64;
2402    let pyr_offset = index_offset + index_len;
2403    let pyr_len = pyramid_meta.len() as u64;
2404    // Optional full-text index between the pyramid and the named graphs.
2405    let text_offset = pyr_offset + pyr_len;
2406    let text_len = text_index.len() as u64;
2407    let named_offset = text_offset + text_len;
2408    let named_len = named_section.len() as u64;
2409    // The quad index sits where the named-graphs section would (the two are
2410    // exclusive).
2411    let quad_len = quad.map_or(0, |q| q.bytes.len() as u64);
2412
2413    // Hash parts in physical order, with the metadata payload prepended when
2414    // present. Omitting it entirely (rather than hashing an empty slice) keeps the
2415    // no-metadata output's hash byte-identical to the pre-metadata writer.
2416    // `verify()` rebuilds this exact list from the header — any section added
2417    // here must be added there too (and covered by a tamper test).
2418    let mut parts: Vec<&[u8]> = Vec::with_capacity(5);
2419    if meta_section_len > 0 {
2420        parts.push(metadata);
2421    }
2422    parts.push(dict_container);
2423    parts.push(&index_container);
2424    parts.push(pyramid_meta);
2425    if text_len > 0 && !text_index_partial {
2426        parts.push(text_index);
2427    }
2428    if named_len > 0 {
2429        parts.push(&named_section);
2430    }
2431
2432    // Length of the trailing schema-pyramid block (0 if none), so a reader can
2433    // fetch just that block for an index/dictionary/summary-free Tier-0 read.
2434    let schema_meta_len = crate::meta::schema_block_len(pyramid_meta);
2435
2436    // The QUAD_INDEX section is not a hash part of its own: the seal inside the
2437    // index container carries its digest, so the content hash covers it — and a
2438    // reader that predates it computes the same hash over the same parts.
2439    let header = Header {
2440        version: crate::header::LEGACY_FORMAT_VERSION,
2441        flags: FLAG_TILE_SYNOPSIS
2442            | if has_quads { FLAG_HAS_QUADS } else { 0 }
2443            | if quad.is_some() {
2444                crate::header::FLAG_QUAD_INDEX
2445            } else {
2446                0
2447            }
2448            | if has_quoted_triples {
2449                FLAG_HAS_QUOTED_TRIPLES
2450            } else {
2451                0
2452            }
2453            | if text_len > 0 && text_index_named {
2454                crate::header::FLAG_TEXT_INDEX_NAMED_GRAPHS
2455            } else {
2456                0
2457            },
2458        metadata_offset: HEADER_LEN as u64,
2459        metadata_len: meta_section_len,
2460        dictionary_offset: dict_offset,
2461        dictionary_len: dict_len,
2462        root_dir_offset: index_offset,
2463        root_dir_len: index_len,
2464        pyramid_meta_offset: if pyr_len > 0 { pyr_offset } else { 0 },
2465        pyramid_meta_len: pyr_len,
2466        dict_codec: codec,
2467        block_codec: codec,
2468        pyramid_levels,
2469        id_widths: 0,
2470        // Every container in the file — default graph and named graphs alike —
2471        // is written from an index built with the same set, so the mask is a
2472        // file-level fact and the default index is its source of truth.
2473        perms: default_index.perms(),
2474        quad_count: default_index.triple_count() as u64 + named_quads,
2475        term_count,
2476        content_hash: content_hash(&parts),
2477        named_graphs_offset: if named_len > 0 { named_offset } else { 0 },
2478        named_graphs_len: named_len,
2479        schema_meta_len,
2480        text_index_offset: if text_len > 0 { text_offset } else { 0 },
2481        text_index_len: text_len,
2482        text_index_partial: text_len > 0 && text_index_partial,
2483        build_info_offset: 0,
2484        build_info_len: 0,
2485        extra_sections: if quad_len > 0 {
2486            vec![crate::header::Section {
2487                kind: crate::header::SectionKind::QuadIndex,
2488                flags: 0,
2489                offset: named_offset,
2490                length: quad_len,
2491            }]
2492        } else {
2493            Vec::new()
2494        },
2495    };
2496
2497    let mut out = Vec::with_capacity(
2498        HEADER_LEN
2499            + metadata.len()
2500            + dict_container.len()
2501            + index_container.len()
2502            + pyramid_meta.len()
2503            + text_index.len()
2504            + named_section.len()
2505            + quad_len as usize
2506            + MAGIC.len(),
2507    );
2508    out.extend_from_slice(&header.to_bytes());
2509    if meta_section_len > 0 {
2510        out.extend_from_slice(metadata);
2511    }
2512    out.extend_from_slice(dict_container);
2513    out.extend_from_slice(&index_container);
2514    out.extend_from_slice(pyramid_meta);
2515    if text_len > 0 {
2516        out.extend_from_slice(text_index);
2517    }
2518    if named_len > 0 {
2519        out.extend_from_slice(&named_section);
2520    }
2521    if let Some(q) = quad {
2522        out.extend_from_slice(&q.bytes);
2523    }
2524    out.extend_from_slice(&MAGIC); // footer marker
2525    out
2526}
2527
2528/// `rdf:type` — the predicate that assigns a class to a resource.
2529pub const RDF_TYPE: &str = "<http://www.w3.org/1999/02/22-rdf-syntax-ns#type>";
2530
2531/// An **ontology-aware** coarse graph: instead of structural communities, group
2532/// entities by their `rdf:type` class and aggregate relations between classes.
2533/// Returns `(subject_class, predicate, object_class, count)` over the default
2534/// graph. Entities with no type are `(untyped)`; literals are `(literal)`.
2535/// `rdf:type` triples themselves define the classes and are not counted as
2536/// relations. This is the dataset's effective schema with instance volumes.
2537pub fn schema_summary(rete: &Rete) -> Vec<(String, String, String, u32)> {
2538    use std::collections::{BTreeMap, HashMap};
2539    let triples = rete.dump(None);
2540
2541    let mut class_of: HashMap<&str, &str> = HashMap::new();
2542    for (s, p, o) in &triples {
2543        if p == RDF_TYPE {
2544            class_of.insert(s.as_str(), o.as_str());
2545        }
2546    }
2547    let classify = |t: &str| -> String {
2548        if let Some(c) = class_of.get(t) {
2549            (*c).to_string()
2550        } else if t.starts_with('"') {
2551            "(literal)".to_string()
2552        } else {
2553            "(untyped)".to_string()
2554        }
2555    };
2556
2557    let mut counts: BTreeMap<(String, String, String), u32> = BTreeMap::new();
2558    for (s, p, o) in &triples {
2559        if p == RDF_TYPE {
2560            continue; // type assertions define classes, not data relations
2561        }
2562        *counts
2563            .entry((classify(s), p.clone(), classify(o)))
2564            .or_default() += 1;
2565    }
2566    counts
2567        .into_iter()
2568        .map(|((a, p, b), c)| (a, p, b, c))
2569        .collect()
2570}
2571
2572/// Class populations: the number of resources of each `rdf:type` class in the
2573/// default graph, descending by count. The instance-count companion to
2574/// [`schema_summary`].
2575pub fn schema_classes(rete: &Rete) -> Vec<(String, u32)> {
2576    use std::collections::BTreeMap;
2577    let mut counts: BTreeMap<String, u32> = BTreeMap::new();
2578    for (_s, p, o) in rete.dump(None) {
2579        if p == RDF_TYPE {
2580            *counts.entry(o).or_default() += 1;
2581        }
2582    }
2583    let mut out: Vec<(String, u32)> = counts.into_iter().collect();
2584    out.sort_by(|a, b| b.1.cmp(&a.1).then_with(|| a.0.cmp(&b.0)));
2585    out
2586}
2587
2588/// Fetch **only** the metadata section (the opaque Dataset Card blob) via a
2589/// [`RangeReader`]: read the 1 KiB header, then the metadata byte range —
2590/// nothing else. This is the index-free CARD tier of the exploration model: a
2591/// remote/S3 client learns the dataset's self-description in **two small range
2592/// requests**, never touching the dictionary, index, or pyramid. Returns `None`
2593/// when the file carries no metadata.
2594///
2595/// Companion to [`Rete::open_ranged`] (which deliberately *skips* the card to
2596/// keep the query path minimal); this is the explicit "I want the card" path.
2597pub fn read_metadata_ranged<R: RangeReader>(reader: &R) -> Result<Option<Vec<u8>>, FileError> {
2598    let head = reader.read_at(0, HEADER_LEN as u64)?;
2599    if crate::v06::open::is_v06(&head) {
2600        let h = crate::v06::Header06::from_bytes(&head)?;
2601        return Ok(crate::v06::open::read_card_build(reader, &h)?.0);
2602    }
2603    let header = Header::from_bytes(&head)?;
2604    if header.metadata_len == 0 {
2605        return Ok(None);
2606    }
2607    let bytes = reader.read_at(header.metadata_offset, header.metadata_len)?;
2608    Ok(Some(bytes))
2609}
2610
2611/// Fetch the metadata (Dataset Card) **and** build-info sections via a
2612/// [`RangeReader`] in the fewest requests: the 1 KiB header, then — because the
2613/// writer lays the build-info section immediately after the metadata — **one**
2614/// coalesced range covering both. A file with only one of the two costs the
2615/// same two requests; a file with neither costs just the header read. This
2616/// keeps "card + build conditions" within the CARD tier's 1 header + 1 range
2617/// budget instead of adding a third request.
2618///
2619/// Returns `(metadata, build_info)`, each `None` when absent.
2620#[allow(clippy::type_complexity)]
2621pub fn read_card_and_build_info_ranged<R: RangeReader>(
2622    reader: &R,
2623) -> Result<(Option<Vec<u8>>, Option<Vec<u8>>), FileError> {
2624    read_card_and_build_info_with_header(reader).map(|(_, m, b)| (m, b))
2625}
2626
2627/// [`read_card_and_build_info_ranged`], also returning the parsed [`Header`] it
2628/// had to read anyway.
2629///
2630/// The header is not a bonus: it carries the content hash and the section
2631/// directory — including whether the file has a TEXT_INDEX section, which the
2632/// card does not state (see
2633/// [`read_text_index_token_table_len_ranged`]). A caller that wants both and
2634/// used the tuple-only form would issue a **second** 1 KiB request for bytes it
2635/// already had, turning the CARD tier's documented "one header + one coalesced
2636/// range" into three requests.
2637#[allow(clippy::type_complexity)]
2638pub fn read_card_and_build_info_with_header<R: RangeReader>(
2639    reader: &R,
2640) -> Result<(Header, Option<Vec<u8>>, Option<Vec<u8>>), FileError> {
2641    let head = reader.read_at(0, HEADER_LEN as u64)?;
2642    let header = Header::from_bytes(&head)?;
2643    if header.version == crate::v06::VERSION {
2644        let h = crate::v06::Header06::from_bytes(&head)?;
2645        let (card, build) = crate::v06::open::read_card_build(reader, &h)?;
2646        return Ok((header, card, build));
2647    }
2648    let meta = (header.metadata_offset, header.metadata_len);
2649    let build = (header.build_info_offset, header.build_info_len);
2650    if meta.1 > 0 && build.1 > 0 && build.0 == meta.0 + meta.1 {
2651        // Adjacent (the layout this crate writes): one read spans both.
2652        let both = reader.read_at(meta.0, meta.1 + build.1)?;
2653        let (m, b) = usize::try_from(meta.1)
2654            .ok()
2655            .and_then(|mid| both.split_at_checked(mid))
2656            .ok_or(FileError::Container("short card read"))?;
2657        return Ok((header, Some(m.to_vec()), Some(b.to_vec())));
2658    }
2659    let fetch = |off: u64, len: u64| -> Result<Option<Vec<u8>>, FileError> {
2660        if len == 0 {
2661            return Ok(None);
2662        }
2663        Ok(Some(reader.read_at(off, len)?))
2664    };
2665    Ok((header, fetch(meta.0, meta.1)?, fetch(build.0, build.1)?))
2666}
2667
2668/// The build-info section of a whole file image, or `None` if absent. The
2669/// payload is an opaque application-layer blob (the CLI stores build-conditions
2670/// JSON there); it is **not** covered by the content hash — see
2671/// [`attach_build_info`].
2672pub fn read_build_info(bytes: &[u8]) -> Result<Option<Vec<u8>>, FileError> {
2673    if crate::v06::open::is_v06(bytes) {
2674        let h = crate::v06::Header06::from_bytes(bytes)?;
2675        let r = crate::reader::SliceReader::new(bytes);
2676        return Ok(crate::v06::open::read_card_build(&r, &h)?.1);
2677    }
2678    let header = Header::from_bytes(bytes)?;
2679    if header.build_info_len == 0 {
2680        return Ok(None);
2681    }
2682    let section = checked_slice(bytes, header.build_info_offset, header.build_info_len)
2683        .ok_or(FileError::Container("build-info section overruns buffer"))?;
2684    Ok(Some(section.to_vec()))
2685}
2686
2687/// Everything needed to splice a build-info section into a finished file
2688/// **without holding the file in memory**: the rewritten header, where the
2689/// section goes, and where the bytes that do not move resume.
2690///
2691/// See [`plan_build_info`] for what the fields mean together: the build-info
2692/// section always follows the metadata section, so the layout is
2693/// `[header][metadata][build info][everything else, verbatim]` — or, for a
2694/// card placed after the data, `[header][data…][metadata][build info][footer]`.
2695#[derive(Debug, Clone, PartialEq, Eq)]
2696pub struct BuildInfoPlan {
2697    /// The header bytes to write at offset 0.
2698    pub header: [u8; HEADER_LEN],
2699    /// Offset the new section is written at: the end of the metadata section
2700    /// (the length of the unchanged prefix up to and including the card).
2701    pub insert: u64,
2702    /// Offset in the **old** file where the bytes that do not move resume. The
2703    /// span `insert..tail_start` is the old build-info section, if any, and is
2704    /// the only part of the old file the new one drops.
2705    pub tail_start: u64,
2706    /// Length of the file this plan produces.
2707    pub new_len: u64,
2708}
2709
2710/// Plan the splice of a build-info section of `info_len` bytes into a file
2711/// whose first [`HEADER_LEN`] bytes are `head` and whose total length is
2712/// `file_len`.
2713///
2714/// The section goes immediately after the metadata section — adjacent, so
2715/// [`read_card_and_build_info_ranged`] fetches card + build info in one
2716/// coalesced range — and every later section's offset shifts by the length
2717/// delta. That shift is the whole of the arithmetic, and it lives here so the
2718/// in-memory splice ([`attach_build_info`]) and any streaming rewriter derive
2719/// the same header from the same rule instead of each carrying a copy that can
2720/// drift.
2721///
2722/// The content hash is **deliberately left untouched**: build info records the
2723/// facts that differ between two builds of identical data (timestamp, builder,
2724/// measured costs), and folding them into the hash would break the
2725/// reproducible-hash property. `verify` accordingly ignores this section, on an
2726/// old reader (which sees an unknown kind-7 entry) and a new one alike — so a
2727/// file that gains, loses or changes a build record **keeps its content
2728/// identity**.
2729///
2730/// `info_len` of 0 plans the removal of an existing section.
2731pub fn plan_build_info(
2732    head: &[u8],
2733    file_len: u64,
2734    info_len: u64,
2735) -> Result<BuildInfoPlan, FileError> {
2736    let mut header = Header::from_bytes(head)?;
2737    // The insert point: immediately after the metadata section, wherever it
2738    // lies — at HEADER_LEN as the in-RAM writers place it, or after the data
2739    // where the memory-bounded build and `recard --in-place` put it
2740    // (`crate::recard::plan_card_append`). No card: right after the header.
2741    let insert = if header.metadata_len > 0 {
2742        header
2743            .metadata_offset
2744            .checked_add(header.metadata_len)
2745            .filter(|&end| header.metadata_offset >= HEADER_LEN as u64 && end <= file_len)
2746            .ok_or(FileError::Container(
2747                "the metadata section lies outside the file",
2748            ))?
2749    } else {
2750        HEADER_LEN as u64
2751    };
2752    let old_len = header.build_info_len;
2753    if old_len > 0 && header.build_info_offset != insert {
2754        return Err(FileError::Container(
2755            "existing build-info section is not adjacent to the metadata",
2756        ));
2757    }
2758    let tail_start = insert.saturating_add(old_len);
2759    if tail_start > file_len || insert > file_len {
2760        return Err(FileError::Container("build-info splice out of bounds"));
2761    }
2762
2763    // Shift every section that lives at or after the old tail.
2764    let shift = |off: &mut u64, len: u64| {
2765        if len > 0 && *off >= tail_start {
2766            *off = *off - old_len + info_len;
2767        }
2768    };
2769    shift(&mut header.dictionary_offset, header.dictionary_len);
2770    shift(&mut header.root_dir_offset, header.root_dir_len);
2771    shift(&mut header.pyramid_meta_offset, header.pyramid_meta_len);
2772    shift(&mut header.named_graphs_offset, header.named_graphs_len);
2773    shift(&mut header.text_index_offset, header.text_index_len);
2774    for s in &mut header.extra_sections {
2775        if s.length > 0 && s.offset >= tail_start {
2776            s.offset = s.offset - old_len + info_len;
2777        }
2778    }
2779    if info_len == 0 {
2780        header.build_info_offset = 0;
2781        header.build_info_len = 0;
2782    } else {
2783        header.build_info_offset = insert;
2784        header.build_info_len = info_len;
2785    }
2786
2787    Ok(BuildInfoPlan {
2788        header: header.to_bytes(),
2789        insert,
2790        tail_start,
2791        new_len: file_len - old_len + info_len,
2792    })
2793}
2794
2795/// Attach (or replace) a **build-info** section in a finished file image,
2796/// returning the new image. Layout and hash semantics: [`plan_build_info`].
2797///
2798/// Passing an empty `info` removes an existing section (or returns the image
2799/// unchanged when there is none).
2800pub fn attach_build_info(image: &[u8], info: &[u8]) -> Result<Vec<u8>, FileError> {
2801    if crate::v06::open::is_v06(image) {
2802        // 0x06: the record lives in the card region; the card stays.
2803        let h = crate::v06::Header06::from_bytes(image)?;
2804        let (card, _) =
2805            crate::v06::open::read_card_build(&crate::reader::SliceReader::new(image), &h)?;
2806        return crate::v06::write::replace_card(
2807            image,
2808            card.as_deref(),
2809            (!info.is_empty()).then_some(info),
2810        );
2811    }
2812    let plan = plan_build_info(image, image.len() as u64, info.len() as u64)?;
2813    let mut out = Vec::with_capacity(plan.new_len as usize);
2814    out.extend_from_slice(&plan.header);
2815    out.extend_from_slice(&image[HEADER_LEN..plan.insert as usize]);
2816    out.extend_from_slice(info);
2817    out.extend_from_slice(&image[plan.tail_start as usize..]);
2818    debug_assert_eq!(out.len() as u64, plan.new_len);
2819    Ok(out)
2820}
2821
2822/// Replace the **metadata** section (the Dataset Card blob) of a finished file
2823/// image, returning the new image. The section is the first one after the
2824/// header, so every later section's offset shifts by the length delta.
2825///
2826/// Unlike [`attach_build_info`], the metadata payload **is** inside the content
2827/// hash — it is part of what the file says about itself — so the hash is
2828/// recomputed here. The result is byte-identical to what the writer would have
2829/// produced had it been handed this payload in the first place
2830/// (`write_dataset_with_metadata`): the section sits at `HEADER_LEN`, every
2831/// later offset is derived from its length, and the hash folds the same parts
2832/// in the same order.
2833///
2834/// Why splice rather than rebuild: the only caller that needs this is a build
2835/// that has already assembled the file and then learned something about it by
2836/// **running** its own starter queries against it. Re-deriving the whole image
2837/// to correct a metadata blob would mean re-parsing the input; a splice costs
2838/// one memcpy and one rehash of an image that is already resident.
2839///
2840/// Passing an empty payload removes the section.
2841pub fn replace_metadata(image: &[u8], metadata: &[u8]) -> Result<Vec<u8>, FileError> {
2842    if crate::v06::open::is_v06(image) {
2843        // 0x06: the card region is rewritten; every section is kept.
2844        let h = crate::v06::Header06::from_bytes(image)?;
2845        let (_, build) =
2846            crate::v06::open::read_card_build(&crate::reader::SliceReader::new(image), &h)?;
2847        return crate::v06::write::replace_card(
2848            image,
2849            (!metadata.is_empty()).then_some(metadata),
2850            build.as_deref(),
2851        );
2852    }
2853    let mut header = Header::from_bytes(image)?;
2854    // The writers in this crate always place metadata (when present) at
2855    // HEADER_LEN; anything else is a file this function must not reshape.
2856    let insert = HEADER_LEN as u64;
2857    if header.metadata_len > 0 && header.metadata_offset != insert {
2858        return Err(FileError::Container(
2859            "metadata section does not sit immediately after the header",
2860        ));
2861    }
2862    let old_len = header.metadata_len;
2863    let tail_start = (insert + old_len) as usize;
2864    if tail_start > image.len() {
2865        return Err(FileError::Container("metadata splice out of bounds"));
2866    }
2867    let new_len = metadata.len() as u64;
2868
2869    let shift = |off: &mut u64, len: u64| {
2870        if len > 0 && *off >= insert + old_len {
2871            *off = *off - old_len + new_len;
2872        }
2873    };
2874    shift(&mut header.dictionary_offset, header.dictionary_len);
2875    shift(&mut header.root_dir_offset, header.root_dir_len);
2876    shift(&mut header.pyramid_meta_offset, header.pyramid_meta_len);
2877    shift(&mut header.named_graphs_offset, header.named_graphs_len);
2878    shift(&mut header.text_index_offset, header.text_index_len);
2879    shift(&mut header.build_info_offset, header.build_info_len);
2880    for s in &mut header.extra_sections {
2881        if s.length > 0 && s.offset >= insert + old_len {
2882            s.offset = s.offset - old_len + new_len;
2883        }
2884    }
2885    header.metadata_offset = insert;
2886    header.metadata_len = new_len;
2887
2888    let mut out = Vec::with_capacity(image.len() - old_len as usize + metadata.len());
2889    out.extend_from_slice(&header.to_bytes());
2890    out.extend_from_slice(metadata);
2891    out.extend_from_slice(&image[tail_start..]);
2892    // The payload changed, so the hash must: recompute it over the same part
2893    // list `verify` will rebuild from this very header, then re-serialize the
2894    // header carrying it.
2895    header.content_hash = hash_of_sections(&header, &out)?;
2896    out[..HEADER_LEN].copy_from_slice(&header.to_bytes());
2897    Ok(out)
2898}
2899
2900/// The content hash of a file image, folded over exactly the parts the writer
2901/// hashed (see `write_dataset_from_parts`): the metadata payload when present,
2902/// then dict, index, pyramid-meta, and — when present — the text index and the
2903/// named graphs. [`verify`] compares this against the header; [`replace_metadata`]
2904/// writes it.
2905fn hash_of_sections(header: &Header, bytes: &[u8]) -> Result<[u8; 16], FileError> {
2906    let slice = |off: u64, len: u64| -> Result<&[u8], FileError> {
2907        // `off` and `len` come STRAIGHT FROM THE FILE HEADER, so a crafted or
2908        // corrupt .rete can set them near u64::MAX and make `off + len` overflow
2909        // — which panics in debug and wraps in release, BEFORE `.get()` gets the
2910        // chance to reject the range. Found by the weekly fuzz run
2911        // (fuzz_targets/open.rs, "attempt to add with overflow"). Saturating the
2912        // sum turns both cases into the same clean Container error.
2913        // `checked_slice` also refuses an offset past this target's usize
2914        // (wasm32), which `as usize` used to wrap onto other bytes.
2915        checked_slice(bytes, off, len).ok_or(FileError::Container("section overruns buffer"))
2916    };
2917    let parts = content_hash_parts(header)
2918        .into_iter()
2919        .map(|(off, len)| slice(off, len))
2920        .collect::<Result<Vec<&[u8]>, FileError>>()?;
2921    Ok(content_hash(&parts))
2922}
2923
2924/// The byte ranges the content hash folds, **in the order it folds them**: the
2925/// metadata payload when present, then dictionary, index, pyramid-meta, and —
2926/// when present — a kind-6 text index and the named graphs. Build info, the
2927/// self-digested kinds (8 LABEL_INDEX, 9 SCHEMA, 11 predicate-restricted text
2928/// index) and any section kind this reader does not know are deliberately
2929/// outside it. The quad index (kind 10) is not a range here either: its digest
2930/// is the seal at the end of the index container, so it is hashed by reference.
2931///
2932/// The one list [`verify`], [`verify_ranged`] and [`replace_metadata`] hash
2933/// over, and what a streaming rewriter (`rete recard`) feeds a hasher while it
2934/// copies — so a section a future format adds to the hash is added here, once,
2935/// and every one of them follows.
2936pub fn content_hash_parts(header: &Header) -> Vec<(u64, u64)> {
2937    let mut parts: Vec<(u64, u64)> = Vec::with_capacity(6);
2938    if header.metadata_len > 0 {
2939        parts.push((header.metadata_offset, header.metadata_len));
2940    }
2941    parts.push((header.dictionary_offset, header.dictionary_len));
2942    parts.push((header.root_dir_offset, header.root_dir_len));
2943    // A zero-length pyramid folds nothing either way; it is listed only when
2944    // present so the ranges a rewriter walks are all real sections.
2945    if header.pyramid_meta_len > 0 {
2946        parts.push((header.pyramid_meta_offset, header.pyramid_meta_len));
2947    }
2948    // A kind-11 (predicate-restricted) text index is self-digested instead.
2949    if header.text_index_len > 0 && !header.text_index_partial {
2950        parts.push((header.text_index_offset, header.text_index_len));
2951    }
2952    if header.named_graphs_len > 0 {
2953        parts.push((header.named_graphs_offset, header.named_graphs_len));
2954    }
2955    parts
2956}
2957
2958/// `verify` for a `0x06` file over any reader: what failed (empty = all good).
2959/// `only` restricts the section checks to one kind (`verify --section`).
2960pub fn verify_ranged_v06(
2961    reader: &dyn crate::reader::RangeReader,
2962    head: &[u8],
2963    only: Option<u8>,
2964) -> Result<Vec<String>, FileError> {
2965    const SLICE: u64 = 8 << 20;
2966    let h = crate::v06::Header06::from_bytes(head)?;
2967    if h.file_len != reader.len() {
2968        return Ok(vec!["file length".to_string()]);
2969    }
2970    let mut bad = Vec::new();
2971    let root_bytes = reader.read_at(crate::v06::HEADER_LEN as u64, h.root_len as u64)?;
2972    if crate::v06::digest16(&root_bytes) != h.root_digest {
2973        bad.push("root".to_string());
2974        return Ok(bad);
2975    }
2976    let root = crate::v06::Root06::decode(&root_bytes, h.internal_codec)?;
2977    if crate::v06::content_hash(&h) != h.content_hash {
2978        bad.push("content hash".to_string());
2979    }
2980    let card = h.card_range();
2981    if only.is_none() && card.len > 0 {
2982        let stored = reader.read_at(card.offset, card.len)?;
2983        if crate::v06::digest16(&stored) != h.card_digest {
2984            bad.push("card".to_string());
2985        }
2986    }
2987    for s in &root.sections {
2988        if only.is_some_and(|k| k != s.kind) {
2989            continue;
2990        }
2991        let a = match s.abs(&h) {
2992            Ok(a) if a.offset + a.len <= h.file_len => a,
2993            _ => {
2994                bad.push(crate::v06::kind::name(s.kind).to_string());
2995                continue;
2996            }
2997        };
2998        let mut hs = blake3::Hasher::new();
2999        let mut at = a.offset;
3000        while at < a.offset + a.len {
3001            let n = (a.offset + a.len - at).min(SLICE);
3002            hs.update(&reader.read_at(at, n)?);
3003            at += n;
3004        }
3005        if hs.finalize().as_bytes()[..16] != s.digest {
3006            bad.push(crate::v06::kind::name(s.kind).to_string());
3007        }
3008    }
3009    if only.is_none() {
3010        let n = crate::v06::FOOTER.len() as u64;
3011        if reader.read_at(h.file_len - n, n)? != crate::v06::FOOTER {
3012            bad.push("footer".to_string());
3013        }
3014    }
3015    Ok(bad)
3016}
3017
3018/// Recompute the content hash from a file image and check it against the header
3019/// — detects corruption or truncation of the payload sections. The
3020/// LABEL_INDEX, SCHEMA and predicate-restricted TEXT_INDEX sections (kinds 8,
3021/// 9, 11) are outside that hash — an older reader's `verify` must keep passing
3022/// on a file that carries them — so each is checked against its own leading
3023/// digest instead. The QUAD_INDEX section (kind 10) is hashed by reference: its
3024/// digest is the seal inside the (hashed) index container, checked here.
3025pub fn verify(bytes: &[u8]) -> Result<bool, FileError> {
3026    if crate::v06::open::is_v06(bytes) {
3027        let r: std::sync::Arc<dyn RangeReader + Send + Sync> =
3028            std::sync::Arc::new(crate::reader::BytesReader(std::sync::Arc::from(bytes)));
3029        return crate::v06::open::V06File::open(r, bytes)?.verify(None);
3030    }
3031    let header = Header::from_bytes(bytes)?;
3032    if hash_of_sections(&header, bytes)? != header.content_hash {
3033        return Ok(false);
3034    }
3035    // The QUAD_INDEX section is covered through the seal's digest (inside the
3036    // hashed index container), so it is checked against that.
3037    if header.has_quad_index() || header.quad_index_section().is_some() {
3038        check_named_layout(&header)?;
3039        let slice = |off: u64, len: u64| -> Result<&[u8], FileError> {
3040            checked_slice(bytes, off, len).ok_or(FileError::Container("section overruns buffer"))
3041        };
3042        let root = slice(header.root_dir_offset, header.root_dir_len)?;
3043        let ranges = container_section_payload_ranges(
3044            root,
3045            header.root_dir_offset,
3046            header.index_container_sections(),
3047        )?;
3048        let seal = ranges
3049            .last()
3050            .ok_or(FileError::Container("missing quad-index seal"))?;
3051        let want = crate::quad::parse_seal(slice(seal.offset, seal.len)?)?;
3052        let q = header
3053            .quad_index_section()
3054            .ok_or(FileError::Container("missing quad-index section"))?;
3055        if crate::quad::digest(slice(q.offset, q.length)?) != want {
3056            return Ok(false);
3057        }
3058    }
3059    for s in self_digested_sections(&header) {
3060        let section = checked_slice(bytes, s.offset, s.length)
3061            .ok_or(FileError::Container("section overruns buffer"))?;
3062        if !crate::label_index::digest_ok(section) {
3063            return Ok(false);
3064        }
3065    }
3066    Ok(true)
3067}
3068
3069/// The sections that carry their own leading 16-byte digest rather than being
3070/// folded into the header content hash: LABEL_INDEX (8), SCHEMA (9) and a
3071/// predicate-restricted TEXT_INDEX (11).
3072fn self_digested_sections(header: &Header) -> Vec<Section> {
3073    [
3074        header.label_index_section(),
3075        header.schema_section(),
3076        header
3077            .section(crate::header::SectionKind::TextIndexPartial)
3078            .filter(|s| s.length > 0),
3079    ]
3080    .into_iter()
3081    .flatten()
3082    .collect()
3083}
3084
3085/// Append LABEL_INDEX / SCHEMA sections (kinds 8, 9) to a finished file image,
3086/// just before its 4-byte footer, and return the new image.
3087///
3088/// They sit outside the content hash, so nothing before them moves and the
3089/// hash is unchanged: only the header's directory gains entries. An empty
3090/// payload is skipped. Each kind may be appended once; a second append of a
3091/// kind the file already carries is an error, as is any other kind.
3092pub fn append_sections(
3093    image: &[u8],
3094    sections: &[(SectionKind, &[u8])],
3095) -> Result<Vec<u8>, FileError> {
3096    let lens: Vec<(SectionKind, u64)> =
3097        sections.iter().map(|(k, b)| (*k, b.len() as u64)).collect();
3098    let (header, body_end) = plan_append(image, image.len() as u64, &lens)?;
3099    let extra: usize = sections.iter().map(|(_, b)| b.len()).sum();
3100    let mut out = Vec::with_capacity(image.len() + extra);
3101    out.extend_from_slice(&header.to_bytes());
3102    out.extend_from_slice(&image[HEADER_LEN..body_end as usize]);
3103    for (_, b) in sections {
3104        out.extend_from_slice(b);
3105    }
3106    out.extend_from_slice(&MAGIC);
3107    Ok(out)
3108}
3109
3110/// The header an append produces, and where the old footer starts.
3111fn plan_append(
3112    head: &[u8],
3113    file_len: u64,
3114    sections: &[(SectionKind, u64)],
3115) -> Result<(Header, u64), FileError> {
3116    let mut header = Header::from_bytes(head)?;
3117    if header.expected_file_len() != Some(file_len) {
3118        return Err(FileError::Container(
3119            "cannot append sections: the file length does not match its header",
3120        ));
3121    }
3122    let body_end = file_len - MAGIC.len() as u64;
3123    let mut at = body_end;
3124    for &(kind, len) in sections {
3125        if !matches!(kind, SectionKind::LabelIndex | SectionKind::Schema) {
3126            return Err(FileError::Container(
3127                "only LABEL_INDEX and SCHEMA sections can be appended",
3128            ));
3129        }
3130        if len == 0 {
3131            continue;
3132        }
3133        if header.section(kind).is_some_and(|s| s.length > 0) {
3134            return Err(FileError::Container(
3135                "the file already carries that section",
3136            ));
3137        }
3138        header = header.with_section(kind, at, len);
3139        at += len;
3140    }
3141    Ok((header, body_end))
3142}
3143
3144/// Open a file's dictionary over a [`RangeReader`] with only its chunk
3145/// directories resident — for the memory-bounded build, which resolves the
3146/// SCHEMA section's few class and predicate ids through the file it just
3147/// wrote.
3148#[cfg(not(target_arch = "wasm32"))]
3149pub(crate) fn open_dictionary_lazy<R: RangeReader + Send + Sync + 'static>(
3150    reader: R,
3151) -> Result<Dictionary, FileError> {
3152    let head = reader.read_at(0, HEADER_LEN as u64)?;
3153    let header = Header::from_bytes(&head)?;
3154    ranged_chunked_dictionary(&std::sync::Arc::new(reader), &header, [true; 4])
3155}
3156
3157/// [`verify`] over a [`RangeReader`](crate::reader::RangeReader), streaming:
3158/// the hashed sections are read in bounded slices and folded into one BLAKE3
3159/// state, so peak memory is one slice, not the file. The same sections, in
3160/// the same order, as `hash_of_sections`; on a 49 GB file the in-memory form
3161/// needs 49 GB.
3162pub fn verify_ranged(reader: &dyn crate::reader::RangeReader) -> Result<bool, FileError> {
3163    /// Read size per step: large enough that a local or remote read is
3164    /// throughput-bound, small enough to be negligible next to any budget.
3165    const SLICE: u64 = 8 << 20;
3166    let len = reader.len();
3167    if len < HEADER_LEN as u64 {
3168        return Err(FileError::Container("file shorter than its header"));
3169    }
3170    let head = reader.read_at(0, HEADER_LEN as u64)?;
3171    if crate::v06::open::is_v06(&head) {
3172        return verify_ranged_v06(reader, &head, None).map(|bad| bad.is_empty());
3173    }
3174    let header = Header::from_bytes(&head)?;
3175    let mut h = blake3::Hasher::new();
3176    for (off, n) in content_hash_parts(&header) {
3177        // Offsets come straight from the header: reject an overrun (or an
3178        // overflowing sum) before reading, as `hash_of_sections` does.
3179        if off.checked_add(n).is_none_or(|end| end > len) {
3180            return Err(FileError::Container("section overruns buffer"));
3181        }
3182        let mut at = off;
3183        while at < off + n {
3184            let take = (off + n - at).min(SLICE);
3185            h.update(&reader.read_at(at, take)?);
3186            at += take;
3187        }
3188    }
3189    let mut out = [0u8; 16];
3190    out.copy_from_slice(&h.finalize().as_bytes()[..16]);
3191    if out != header.content_hash {
3192        return Ok(false);
3193    }
3194    // The QUAD_INDEX section, against the digest in the index container's seal
3195    // (see `verify`), streamed the same way.
3196    if header.has_quad_index() || header.quad_index_section().is_some() {
3197        check_named_layout(&header)?;
3198        let sections = header.index_container_sections();
3199        let seal = locate_container_section_ranged(
3200            reader,
3201            header.root_dir_offset,
3202            header.root_dir_len,
3203            sections - 1,
3204            sections as u64,
3205        )?;
3206        let want = crate::quad::parse_seal(&reader.read_at(seal.offset, seal.len)?)?;
3207        let q = header
3208            .quad_index_section()
3209            .ok_or(FileError::Container("missing quad-index section"))?;
3210        if q.offset.checked_add(q.length).is_none_or(|end| end > len) {
3211            return Err(FileError::Container("section overruns buffer"));
3212        }
3213        let mut h = blake3::Hasher::new();
3214        let mut at = q.offset;
3215        while at < q.offset + q.length {
3216            let take = (q.offset + q.length - at).min(SLICE);
3217            h.update(&reader.read_at(at, take)?);
3218            at += take;
3219        }
3220        let mut got = [0u8; 16];
3221        got.copy_from_slice(&h.finalize().as_bytes()[..16]);
3222        if got != want {
3223            return Ok(false);
3224        }
3225    }
3226    // The self-digested sections (kinds 8, 9, 11), streamed the same way.
3227    for s in self_digested_sections(&header) {
3228        if s.length < 16 || s.offset.checked_add(s.length).is_none_or(|end| end > len) {
3229            return Err(FileError::Container("section overruns buffer"));
3230        }
3231        let want = reader.read_at(s.offset, 16)?;
3232        let mut h = blake3::Hasher::new();
3233        let mut at = s.offset + 16;
3234        while at < s.offset + s.length {
3235            let take = (s.offset + s.length - at).min(SLICE);
3236            h.update(&reader.read_at(at, take)?);
3237            at += take;
3238        }
3239        if h.finalize().as_bytes()[..16] != want[..] {
3240            return Ok(false);
3241        }
3242    }
3243    Ok(true)
3244}
3245
3246/// Faults the pyramid meta in on first access. `None` = the fetch failed.
3247type PyramidLoader = Box<dyn Fn() -> Option<PyramidMeta> + Send + Sync>;
3248
3249/// The pyramid-meta section, held either resident (eager opens) or deferred
3250/// (the lazy remote open). SPARQL never touches the pyramid, but on a Wikidata
3251/// file it can be tens of MB (114k communities, millions of superedges), so a
3252/// remote SPARQL query must not pay to fetch it — it faults in only when a
3253/// community/pyramid query actually calls [`Rete::pyramid`].
3254enum PyramidSlot {
3255    Resident(Option<PyramidMeta>),
3256    Lazy {
3257        loader: PyramidLoader,
3258        cell: std::sync::OnceLock<Option<PyramidMeta>>,
3259    },
3260}
3261
3262/// A section held resident (eager opens) or faulted in on first use (lazy
3263/// opens) — the LABEL_INDEX and SCHEMA sections. `None` inside means the file
3264/// has no such section, or it could not be read or parsed: both are optional
3265/// and derived, so a bad one degrades to "absent" (and `verify` reports it).
3266pub(crate) enum SectionSlot<T> {
3267    Resident(Option<T>),
3268    Lazy {
3269        loader: Box<dyn Fn() -> Option<T> + Send + Sync>,
3270        cell: std::sync::OnceLock<Option<T>>,
3271    },
3272}
3273
3274impl<T> SectionSlot<T> {
3275    /// The section, faulting it in on the lazy path.
3276    pub(crate) fn get(&self) -> Option<&T> {
3277        match self {
3278            SectionSlot::Resident(v) => v.as_ref(),
3279            SectionSlot::Lazy { loader, cell } => cell.get_or_init(loader).as_ref(),
3280        }
3281    }
3282
3283    /// The section only if it is resident or already faulted — never a read.
3284    pub(crate) fn get_if_loaded(&self) -> Option<&T> {
3285        match self {
3286            SectionSlot::Resident(v) => v.as_ref(),
3287            SectionSlot::Lazy { cell, .. } => cell.get().and_then(|o| o.as_ref()),
3288        }
3289    }
3290}
3291
3292/// The LABEL_INDEX of a file image held in memory.
3293fn resident_label_slot<'a>(
3294    header: &Header,
3295    region: &dyn Fn(u64, u64) -> Result<&'a [u8], FileError>,
3296) -> SectionSlot<crate::label_index::LabelIndex> {
3297    SectionSlot::Resident(header.label_index_section().and_then(|s| {
3298        let bytes = region(s.offset, s.length).ok()?;
3299        crate::label_index::LabelIndex::from_section(bytes, header.block_codec).ok()
3300    }))
3301}
3302
3303/// The SCHEMA section of a file image held in memory.
3304fn resident_schema_slot<'a>(
3305    header: &Header,
3306    region: &dyn Fn(u64, u64) -> Result<&'a [u8], FileError>,
3307) -> SectionSlot<crate::schema_section::SchemaMeta> {
3308    SectionSlot::Resident(header.schema_section().and_then(|s| {
3309        crate::schema_section::SchemaMeta::decode(region(s.offset, s.length).ok()?).ok()
3310    }))
3311}
3312
3313/// The deferred LABEL_INDEX of a ranged open: the first search reads the
3314/// section head (preamble + root, ≤ [`crate::label_index::HEAD_READ`] bytes),
3315/// then one directory node per level and the data blocks it needs.
3316fn ranged_label_slot<R: RangeReader + Send + Sync + 'static>(
3317    reader: &std::sync::Arc<R>,
3318    header: &Header,
3319) -> SectionSlot<crate::label_index::LabelIndex> {
3320    let Some(s) = header.label_index_section() else {
3321        return SectionSlot::Resident(None);
3322    };
3323    let r = reader.clone();
3324    let codec = header.block_codec;
3325    SectionSlot::Lazy {
3326        loader: Box::new(move || {
3327            let head = r
3328                .read_at(s.offset, crate::label_index::HEAD_READ.min(s.length))
3329                .ok()?;
3330            let rr = r.clone();
3331            let loader = Box::new(move |off: u64, len: u64| rr.read_at(s.offset + off, len).ok());
3332            crate::label_index::LabelIndex::from_head(&head, s.length, codec, loader).ok()
3333        }),
3334        cell: std::sync::OnceLock::new(),
3335    }
3336}
3337
3338/// The deferred SCHEMA section of a ranged open: one range read, on first use.
3339fn ranged_schema_slot<R: RangeReader + Send + Sync + 'static>(
3340    reader: &std::sync::Arc<R>,
3341    header: &Header,
3342) -> SectionSlot<crate::schema_section::SchemaMeta> {
3343    let Some(s) = header.schema_section() else {
3344        return SectionSlot::Resident(None);
3345    };
3346    let r = reader.clone();
3347    SectionSlot::Lazy {
3348        loader: Box::new(move || {
3349            let bytes = r.read_at(s.offset, s.length).ok()?;
3350            crate::schema_section::SchemaMeta::decode(&bytes).ok()
3351        }),
3352        cell: std::sync::OnceLock::new(),
3353    }
3354}
3355
3356/// Read and decode a file's SCHEMA section over a [`RangeReader`] — the
3357/// header and one range. `Ok(None)` when the file has none (or it is
3358/// malformed: the section is optional and derived).
3359pub fn read_schema_ranged<R: RangeReader + ?Sized>(
3360    reader: &R,
3361) -> Result<Option<crate::schema_section::SchemaMeta>, FileError> {
3362    let head = reader.read_at(0, HEADER_LEN as u64)?;
3363    let header = Header::from_bytes(&head)?;
3364    Ok(read_schema_with_header(reader, &header))
3365}
3366
3367fn read_schema_with_header<R: RangeReader + ?Sized>(
3368    reader: &R,
3369    header: &Header,
3370) -> Option<crate::schema_section::SchemaMeta> {
3371    if header.version == crate::v06::VERSION {
3372        let b = crate::v06::open::read_schema_bytes(reader).ok()??;
3373        return crate::schema_section::SchemaMeta::decode(&b).ok();
3374    }
3375    let s = header.schema_section()?;
3376    let bytes = reader.read_at(s.offset, s.length).ok()?;
3377    crate::schema_section::SchemaMeta::decode(&bytes).ok()
3378}
3379
3380/// Faults the text index in on first search. `None` = the file has none, or the
3381/// fetch/parse failed.
3382type TextIndexLoader = Box<dyn Fn() -> Option<crate::text_index::TextIndex> + Send + Sync>;
3383
3384/// Answers the token table's byte length from the section's first ≤10 bytes —
3385/// the length varint alone, never the table it measures.
3386type TokenTableProbe = Box<dyn Fn() -> Option<u64> + Send + Sync>;
3387
3388/// The TEXT_INDEX section, held either resident (eager opens decode the whole
3389/// thing) or deferred (the lazy remote open keeps only a loader that fetches the
3390/// token table on first search, then faults posting lists one at a time). SPARQL
3391/// never touches it, so the lazy remote path keeps its small range budget.
3392///
3393/// Both variants can also answer how long the section's leading **token table**
3394/// is without faulting it — the figure a caller needs to state what a first
3395/// search will cost, since `header.text_index_len` counts the postings blob too
3396/// and overstates it by several times on a literal-heavy graph.
3397enum TextIndexSlot {
3398    Resident {
3399        index: Option<crate::text_index::TextIndex>,
3400        /// Measured from the section bytes this open already held; `None` when
3401        /// the file carries no text index.
3402        token_table_len: Option<u64>,
3403    },
3404    Lazy {
3405        loader: TextIndexLoader,
3406        cell: std::sync::OnceLock<Option<crate::text_index::TextIndex>>,
3407        /// One ≤10-byte range read, memoized in `token_table_cell`.
3408        token_table: TokenTableProbe,
3409        token_table_cell: std::sync::OnceLock<Option<u64>>,
3410    },
3411}
3412
3413impl TextIndexSlot {
3414    /// The parsed index, faulting it in (token table first) on the lazy path.
3415    fn index(&self) -> Option<&crate::text_index::TextIndex> {
3416        match self {
3417            TextIndexSlot::Resident { index, .. } => index.as_ref(),
3418            TextIndexSlot::Lazy { loader, cell, .. } => cell.get_or_init(loader).as_ref(),
3419        }
3420    }
3421
3422    /// Byte length of the section's leading token table — see
3423    /// [`Rete::text_index_token_table_len`].
3424    fn token_table_len(&self) -> Option<u64> {
3425        match self {
3426            TextIndexSlot::Resident {
3427                token_table_len, ..
3428            } => *token_table_len,
3429            TextIndexSlot::Lazy {
3430                token_table,
3431                token_table_cell,
3432                ..
3433            } => *token_table_cell.get_or_init(token_table),
3434        }
3435    }
3436}
3437
3438/// Decode a whole TEXT_INDEX section held in memory, measuring its token table
3439/// on the way past. `section` is `None` when the file carries no text index.
3440fn resident_text_index_slot(section: Option<&[u8]>, codec: u8) -> Result<TextIndexSlot, FileError> {
3441    let Some(section) = section else {
3442        return Ok(TextIndexSlot::Resident {
3443            index: None,
3444            token_table_len: None,
3445        });
3446    };
3447    Ok(TextIndexSlot::Resident {
3448        index: Some(
3449            crate::text_index::TextIndex::from_section(section, codec)
3450                .map_err(|_| FileError::Container("malformed text index"))?,
3451        ),
3452        token_table_len: crate::text_index::TextIndex::postings_base(section).map(|b| b as u64),
3453    })
3454}
3455
3456/// Byte length of the TEXT_INDEX section's leading **token table**, from a
3457/// header a caller has already parsed — the prefix range a first
3458/// [`Rete::text_search`] fetches, and the honest figure to quote as its cost
3459/// (`header.text_index_len` counts the postings blob too and overstates it
3460/// several-fold). `None` when the file carries no text index, or its first
3461/// bytes could not be read.
3462///
3463/// Costs **one range read of ≤10 bytes** — the section's leading length varint
3464/// and nothing else, never the table it measures. This is the free-standing
3465/// companion to [`Rete::text_index_token_table_len`], for readers that hold a
3466/// [`Header`] and a [`RangeReader`] but never open the file (the CARD tier:
3467/// `rete card`, `rete card-url`).
3468pub fn read_text_index_token_table_len_ranged<R: RangeReader + ?Sized>(
3469    reader: &R,
3470    header: &Header,
3471) -> Option<u64> {
3472    if header.version == crate::v06::VERSION {
3473        return crate::v06::open::read_token_table_bytes(reader);
3474    }
3475    let (ti_off, ti_len) = header.text_index_payload();
3476    if ti_len == 0 {
3477        return None;
3478    }
3479    read_token_table_len(reader, ti_off, ti_len)
3480}
3481
3482/// Byte length of the TEXT_INDEX section's leading token table — its length
3483/// varint plus the compressed table, i.e. the prefix range a first search
3484/// fetches. Reads the section's first ≤10 bytes and nothing else.
3485fn read_token_table_len<R: RangeReader + ?Sized>(reader: &R, off: u64, len: u64) -> Option<u64> {
3486    let head = reader.read_at(off, 10u64.min(len)).ok()?;
3487    let (ttlen, n) = crate::varint::read_uvarint(&head)?;
3488    Some((n as u64 + ttlen).min(len))
3489}
3490
3491/// A named graph's container decodes RESIDENT (one range read, one decode) up
3492/// to this size on the lazy path; a larger container opens as a remote-lazy
3493/// tile index instead, so one selective `GRAPH <g>` query over a multi-GB
3494/// graph faults only the tile directories plus the tiles it touches.
3495const NAMED_GRAPH_RESIDENT_MAX: u64 = 1 << 20; // 1 MiB
3496
3497/// Largest quad-index graph (in quads) a many-graph walk evaluates through an
3498/// owned, dropped view ([`Rete::named_graph_transient`]) rather than a memoised
3499/// one. A view is ~1–2 KB of descriptors whatever the graph's size, so keeping
3500/// one per graph is what grows with the graph count; past this size a graph is
3501/// worth memoising, and its rows are no longer worth materializing per graph.
3502pub const TRANSIENT_GRAPH_QUADS: u64 = 1024;
3503
3504/// Entries per lazily-allocated named-graph slab. The outer table is sized by
3505/// the section's leading count varint, which is UNTRUSTED — slabbing keeps a
3506/// hostile count from ballooning the allocation (the outer table costs a few
3507/// bytes per claimed slab; entry storage only materializes when touched).
3508const NAMED_SLAB: usize = 1024;
3509
3510/// First read of the named-graph walk, and the floor its ramp resets to. One
3511/// chunk covers a few dozen small-graph records, so a `LIMIT`ed `GRAPH ?g`
3512/// pays exactly one small read — the targeted-query win this size protects.
3513const NAMED_WALK_CHUNK: u64 = 64 * 1024;
3514
3515/// Ceiling of the walk's geometric read ramp — the read size an exhaustive
3516/// walk converges to (and starts at, when the demand is known exhaustive).
3517///
3518/// Sized from measurement, not taste. Against the nkod reference file
3519/// (67.9 MB, 31,974 named graphs, 65.4 MB section) the fixed 64 KiB walk cost
3520/// 262 requests / ~16.5 s in Chromium where the old eager whole-section fetch
3521/// cost ~8.5 s — i.e. ~31 ms of per-request overhead on ~8 s of transfer
3522/// (~7.7 MB/s). At 8 MiB per read a chunk transfers in ~1 s, putting the
3523/// request overhead near 3% of total time; halving the cap to 4 MiB doubles
3524/// that overhead, doubling the cap to 16 MiB buys back under 2% while doubling
3525/// both the walk buffer a wasm32 session must hold resident and the worst-case
3526/// over-read past the point a query stops. 8 MiB is the knee.
3527const NAMED_WALK_CHUNK_MAX: u64 = 8 * 1024 * 1024;
3528
3529/// One named graph, walked lazily out of the NAMED_GRAPHS section.
3530#[derive(Default)]
3531struct NamedEntry {
3532    /// `(iri, absolute byte range of the graph's index container)` — filled
3533    /// when the sequential directory walk reaches this slot.
3534    meta: std::sync::OnceLock<(String, ByteRange)>,
3535    /// The graph's permutation index, opened on first access. Boxed: most
3536    /// entries of a many-graph dataset are never touched by a given query.
3537    index: std::sync::OnceLock<Box<GraphIndex>>,
3538}
3539
3540/// Sequential position of the lazy named-graphs walk. The section has no
3541/// random-access directory — each entry's offset is the previous entry's end —
3542/// so entries are parsed strictly in order and memoised.
3543#[derive(Default)]
3544struct NamedWalk {
3545    /// Next entry index to parse.
3546    next: usize,
3547    /// Absolute file offset of that entry's header; `0` = the leading count
3548    /// varint has not been consumed yet (offset 0 is the file header, never
3549    /// inside a section).
3550    pos: u64,
3551    /// Read buffer carried ACROSS walk calls: `buf` holds the section bytes at
3552    /// `[buf_off, buf_off + buf.len())`. Entry-by-entry iteration (the
3553    /// `GRAPH ?g` path advances one entry per call) parses ~30 small-graph
3554    /// headers per 64 KiB chunk instead of re-reading a chunk per entry.
3555    /// [`LazyNamedGraphs::open_graph`] also serves whole small containers out
3556    /// of it, so a bulk chunk's ride-along payloads are never re-fetched.
3557    buf: Vec<u8>,
3558    buf_off: u64,
3559    /// Current read size of the geometric ramp (`0` = not started; treated as
3560    /// [`NAMED_WALK_CHUNK`]). Doubles on every buffer refill up to
3561    /// [`NAMED_WALK_CHUNK_MAX`], so a walk that keeps going converges to bulk
3562    /// reads after fetching at most as many bytes as it already used (the
3563    /// doubling bound: wasted read-ahead ≤ useful bytes). Reset to the floor
3564    /// whenever an oversized container is parsed — see `big_seen`.
3565    chunk: u64,
3566    /// A container larger than [`NAMED_GRAPH_RESIDENT_MAX`] has been parsed.
3567    /// Such graphs open tile-lazily (their payload bytes are NOT wanted by the
3568    /// walk), so ride-along prefetch stops paying off: the ramp resets and the
3569    /// exhaustive-demand hint stops seeding full-size reads. Without this, a
3570    /// header walk over a file of multi-MB containers would fetch the payloads
3571    /// it exists to hop over.
3572    big_seen: bool,
3573}
3574
3575/// The NAMED_GRAPHS section opened lazily over a [`RangeReader`]: nothing is
3576/// fetched at open. The section is a count varint followed by
3577/// `iri_len | iri | container_len | container` records, so the directory is
3578/// walked forward on demand — header bytes only, skipping container payloads —
3579/// and each touched graph's index is decoded (or opened tile-lazily) on first
3580/// access. A `LIMIT`ed `GRAPH ?g` query therefore reads a prefix of the
3581/// section; only a query that genuinely touches every graph (e.g. a full
3582/// `COUNT` over `GRAPH ?g`) walks it all — which is exactly what the eager
3583/// open used to fetch unconditionally.
3584///
3585/// **Failure contract** (same as lazy tiles/chunks): a failed read or a
3586/// malformed record sets a sticky flag and the accessor returns `None`;
3587/// nothing failed is ever memoised, so a later evaluation retries.
3588struct LazyNamedGraphs {
3589    reader: std::sync::Arc<dyn RangeReader + Send + Sync>,
3590    /// Absolute byte range of the NAMED_GRAPHS section.
3591    section: ByteRange,
3592    codec: u8,
3593    has_synopsis: bool,
3594    read_concurrency: usize,
3595    /// The file's permutation set — every graph container in a file carries
3596    /// the same one, so it comes from the header rather than per graph.
3597    perms: PermSet,
3598    /// `(count, slab table)` — set once the leading count varint is read.
3599    /// Slabs allocate on first touch; boxed slices never move, so `&` handed
3600    /// out to entries stay valid for `&self`'s lifetime with no unsafe.
3601    #[allow(clippy::type_complexity)]
3602    dir: std::sync::OnceLock<(usize, Box<[std::sync::OnceLock<Box<[NamedEntry]>>]>)>,
3603    walk: std::sync::Mutex<NamedWalk>,
3604    failed: std::sync::atomic::AtomicBool,
3605    /// Tile-cache byte cap applied to each LARGE (lazily-opened, loader-bearing)
3606    /// named graph as it is opened (bounded-export phase 2). `u64::MAX` =
3607    /// unlimited. Small resident graphs (no loader, ≤ 1 MiB) are never capped —
3608    /// an evicted tile could not be reloaded — but they are bounded by size and
3609    /// released per slot, so they need no cap.
3610    tile_cap: std::sync::atomic::AtomicU64,
3611}
3612
3613impl LazyNamedGraphs {
3614    fn new(
3615        reader: std::sync::Arc<dyn RangeReader + Send + Sync>,
3616        section: ByteRange,
3617        codec: u8,
3618        has_synopsis: bool,
3619        read_concurrency: usize,
3620        perms: PermSet,
3621    ) -> Self {
3622        LazyNamedGraphs {
3623            reader,
3624            section,
3625            codec,
3626            has_synopsis,
3627            read_concurrency,
3628            perms,
3629            dir: std::sync::OnceLock::new(),
3630            walk: std::sync::Mutex::new(NamedWalk::default()),
3631            failed: std::sync::atomic::AtomicBool::new(false),
3632            tile_cap: std::sync::atomic::AtomicU64::new(u64::MAX),
3633        }
3634    }
3635
3636    /// Set the tile-cache cap for large named graphs: records it for graphs
3637    /// opened later, and re-caps every large graph already opened. Small
3638    /// resident (loaderless) graphs are left unlimited.
3639    fn set_tile_cap(&self, cap: u64) {
3640        self.tile_cap
3641            .store(cap, std::sync::atomic::Ordering::Relaxed);
3642        self.for_each_opened(|g| {
3643            if g.is_lazy() {
3644                g.set_cache_cap(cap);
3645            }
3646        });
3647    }
3648
3649    /// Record a fetch/parse failure (checked by `index_incomplete`).
3650    fn fail(&self) {
3651        self.failed
3652            .store(true, std::sync::atomic::Ordering::Relaxed);
3653    }
3654
3655    /// The directory scaffold, reading the leading count varint on first call.
3656    #[allow(clippy::type_complexity)]
3657    fn directory(&self) -> Option<&(usize, Box<[std::sync::OnceLock<Box<[NamedEntry]>>]>)> {
3658        if let Some(d) = self.dir.get() {
3659            return Some(d);
3660        }
3661        let end = self.section.offset.checked_add(self.section.len)?;
3662        let (n, used) = match read_uvarint_at(&self.reader, self.section.offset, end) {
3663            Ok(v) => v,
3664            Err(_) => {
3665                self.fail();
3666                return None;
3667            }
3668        };
3669        // Each record needs at least 2 bytes of framing (two length varints),
3670        // so a count beyond len/2 is malformed — refuse before sizing by it.
3671        if n > self.section.len / 2 {
3672            self.fail();
3673            return None;
3674        }
3675        let Ok(n) = usize::try_from(n) else {
3676            self.fail();
3677            return None;
3678        };
3679        let slabs = n.div_ceil(NAMED_SLAB);
3680        let table: Box<[std::sync::OnceLock<Box<[NamedEntry]>>]> =
3681            (0..slabs).map(|_| std::sync::OnceLock::new()).collect();
3682        let _ = self.dir.set((n, table));
3683        {
3684            // Idempotent under an init race: both racers computed the same
3685            // `used` from the same bytes.
3686            let mut w = self.walk.lock().unwrap();
3687            if w.pos == 0 {
3688                w.pos = self.section.offset + used;
3689            }
3690        }
3691        self.dir.get()
3692    }
3693
3694    fn count(&self) -> usize {
3695        self.directory().map(|(n, _)| *n).unwrap_or(0)
3696    }
3697
3698    /// The `i`-th entry's cell (allocating its slab on first touch).
3699    fn entry(&self, i: usize) -> Option<&NamedEntry> {
3700        let (n, table) = self.directory()?;
3701        if i >= *n {
3702            return None;
3703        }
3704        let slab = table[i / NAMED_SLAB].get_or_init(|| {
3705            let len = NAMED_SLAB.min(n - (i / NAMED_SLAB) * NAMED_SLAB);
3706            (0..len).map(|_| NamedEntry::default()).collect()
3707        });
3708        slab.get(i % NAMED_SLAB)
3709    }
3710
3711    /// The next walk read's length, advancing the geometric ramp: at least
3712    /// `need` (a long IRI's header), at most the section remainder, and
3713    /// otherwise the current ramp value — which doubles per refill up to
3714    /// [`NAMED_WALK_CHUNK_MAX`]. The doubling keeps a stopping walk's wasted
3715    /// read-ahead bounded by the bytes it already used, while a walk that
3716    /// keeps going converges to bulk reads: the runtime safety net for
3717    /// exhaustive-looking queries the demand analysis could not prove.
3718    fn next_read_len(w: &mut NamedWalk, pos: u64, end: u64, need: u64) -> u64 {
3719        let cur = w.chunk.max(NAMED_WALK_CHUNK);
3720        w.chunk = (cur * 2).min(NAMED_WALK_CHUNK_MAX);
3721        cur.max(need).min(end - pos)
3722    }
3723
3724    /// Walk entry headers forward until `entries[upto].meta` is set. Reads the
3725    /// section in chunks and hops over container payloads, so the walk fetches
3726    /// O(headers) — not O(section) — where containers are large; small
3727    /// containers ride along in the same chunk and cost nothing extra.
3728    ///
3729    /// `exhaustive` is the demand hint from the query engine: the caller has
3730    /// PROVEN it will visit and open every named graph (an unrestricted
3731    /// `GRAPH ?g` under a consumer that drains the pipeline), so the walk
3732    /// starts at full-size reads instead of ramping up to them — the whole
3733    /// section arrives in `section/8 MiB` requests instead of hundreds. The
3734    /// hint is per-call, never memoised: a later targeted query on the same
3735    /// handle walks small again. It also defers to the layout: once an
3736    /// oversized (tile-lazy) container is seen, ride-along bytes stop being
3737    /// useful and the hint no longer seeds bulk reads.
3738    fn ensure_meta(&self, upto: usize, exhaustive: bool) -> Option<()> {
3739        let target = self.entry(upto)?;
3740        if target.meta.get().is_some() {
3741            return Some(());
3742        }
3743        let end = self.section.offset.checked_add(self.section.len)?;
3744        let mut w = self.walk.lock().unwrap();
3745        if exhaustive && !w.big_seen {
3746            w.chunk = NAMED_WALK_CHUNK_MAX;
3747        }
3748        while w.next <= upto {
3749            let pos = w.pos;
3750            if pos >= end {
3751                // The count varint promised more entries than the bytes hold.
3752                self.fail();
3753                return None;
3754            }
3755            // Make sure the whole header (two varints + the IRI, ≤ 20+iri_len
3756            // bytes) is buffered; refetch a larger window when the IRI is long.
3757            let have =
3758                |b: &[u8], off: u64, need: u64| pos >= off && pos + need <= off + b.len() as u64;
3759            if !have(&w.buf, w.buf_off, 20.min(end - pos)) {
3760                let len = Self::next_read_len(&mut w, pos, end, 20.min(end - pos));
3761                w.buf = match self.reader.read_at(pos, len) {
3762                    Ok(b) => b,
3763                    Err(_) => {
3764                        self.fail();
3765                        return None;
3766                    }
3767                };
3768                w.buf_off = pos;
3769            }
3770            // Every slice of `buf` below is checked: a reader may return fewer
3771            // bytes than asked, and the lengths are untrusted.
3772            let rel = (pos - w.buf_off) as usize;
3773            let Some((ilen, u1)) = w.buf.get(rel..).and_then(read_uvarint) else {
3774                self.fail();
3775                return None;
3776            };
3777            let Some(iri_end) = (u1 as u64)
3778                .checked_add(ilen)
3779                .and_then(|n| pos.checked_add(n))
3780                .filter(|&e| e <= end)
3781            else {
3782                self.fail(); // IRI overruns the section: malformed
3783                return None;
3784            };
3785            // iri_len + iri + container_len varint
3786            let header_need = (iri_end - pos).saturating_add(10);
3787            if !have(&w.buf, w.buf_off, header_need.min(end - pos)) {
3788                let len = Self::next_read_len(&mut w, pos, end, header_need);
3789                w.buf = match self.reader.read_at(pos, len) {
3790                    Ok(b) => b,
3791                    Err(_) => {
3792                        self.fail();
3793                        return None;
3794                    }
3795                };
3796                w.buf_off = pos;
3797            }
3798            let rel = (pos - w.buf_off) as usize;
3799            let istart = rel + u1;
3800            let Some(iend) = usize::try_from(ilen)
3801                .ok()
3802                .and_then(|l| istart.checked_add(l))
3803                .filter(|&e| e <= w.buf.len())
3804            else {
3805                self.fail(); // short read: the header is not all here
3806                return None;
3807            };
3808            let iri = String::from_utf8_lossy(&w.buf[istart..iend]).into_owned();
3809            let Some((clen, u2)) = read_uvarint(&w.buf[iend..]) else {
3810                self.fail();
3811                return None;
3812            };
3813            let cend = match iri_end
3814                .checked_add(u2 as u64)
3815                .and_then(|cstart| cstart.checked_add(clen))
3816            {
3817                Some(e) if e <= end => e,
3818                _ => {
3819                    self.fail(); // container overruns the section: malformed
3820                    return None;
3821                }
3822            };
3823            if clen > NAMED_GRAPH_RESIDENT_MAX {
3824                // This graph opens tile-lazily: the walk exists to HOP its
3825                // payload, so big read-ahead here is pure over-fetch. Drop the
3826                // ramp back to the floor and stop the exhaustive hint from
3827                // re-seeding it (headers stay cheap on payload-heavy files).
3828                w.big_seen = true;
3829                w.chunk = NAMED_WALK_CHUNK;
3830            }
3831            let range = ByteRange {
3832                offset: cend - clen,
3833                len: clen,
3834            };
3835            if let Some(e) = self.entry(w.next) {
3836                let _ = e.meta.set((iri, range));
3837            }
3838            w.pos = cend;
3839            w.next += 1;
3840        }
3841        Some(())
3842    }
3843
3844    /// The `i`-th graph's IRI — a header walk, never an index decode.
3845    /// `exhaustive`: see [`ensure_meta`](Self::ensure_meta).
3846    fn name_at(&self, i: usize, exhaustive: bool) -> Option<&str> {
3847        self.ensure_meta(i, exhaustive)?;
3848        self.entry(i)?.meta.get().map(|(iri, _)| iri.as_str())
3849    }
3850
3851    /// The `i`-th graph as `(iri, index)`, opening the index on first access.
3852    /// `exhaustive`: see [`ensure_meta`](Self::ensure_meta).
3853    fn graph_at(&self, i: usize, exhaustive: bool) -> Option<(&str, &GraphIndex)> {
3854        self.ensure_meta(i, exhaustive)?;
3855        let e = self.entry(i)?;
3856        let (iri, range) = e.meta.get()?;
3857        if e.index.get().is_none() {
3858            let opened = self.open_graph(*range)?;
3859            let _ = e.index.set(Box::new(opened));
3860        }
3861        Some((iri.as_str(), e.index.get()?.as_ref()))
3862    }
3863
3864    /// A small container's bytes: out of the walk buffer when the last chunk
3865    /// already carried them (bulk chunks always do — that is their point),
3866    /// else one targeted range read. Never called for tile-lazy containers.
3867    fn container_bytes(&self, range: ByteRange) -> Option<Vec<u8>> {
3868        {
3869            let w = self.walk.lock().unwrap();
3870            let buf_end = w.buf_off + w.buf.len() as u64;
3871            if range.offset >= w.buf_off && range.offset + range.len <= buf_end {
3872                let a = (range.offset - w.buf_off) as usize;
3873                return Some(w.buf[a..a + range.len as usize].to_vec());
3874            }
3875        }
3876        match self.reader.read_at(range.offset, range.len) {
3877            // A short read is a failed read, not a shorter container.
3878            Ok(b) if b.len() as u64 == range.len => Some(b),
3879            _ => {
3880                self.fail();
3881                None
3882            }
3883        }
3884    }
3885
3886    /// Open one graph's index container: small ones resident (one read),
3887    /// large ones as a remote-lazy tile index — the same machinery as the
3888    /// default graph, so a selective query over a huge named graph fetches
3889    /// only the tile directories plus the tiles it touches.
3890    fn open_graph(&self, range: ByteRange) -> Option<GraphIndex> {
3891        if range.len <= NAMED_GRAPH_RESIDENT_MAX {
3892            let bytes = self.container_bytes(range)?;
3893            match decode_index_container(&bytes, self.codec, self.perms) {
3894                Ok(g) => {
3895                    // Resident, so counting is a pass over tile headers: the
3896                    // planner sizes this graph's patterns by it.
3897                    g.set_size_hint(u64::from(g.triple_count()));
3898                    Some(g)
3899                }
3900                Err(_) => {
3901                    self.fail();
3902                    None
3903                }
3904            }
3905        } else {
3906            match open_index_container_lazy(
3907                &self.reader,
3908                range,
3909                self.codec,
3910                self.has_synopsis,
3911                self.read_concurrency,
3912                self.perms,
3913                self.perms.len(),
3914            ) {
3915                Ok((g, _, _)) => {
3916                    // A large graph faults its tiles through a loader, so bound
3917                    // its tile residency to the budget's share (an evicted tile
3918                    // simply re-faults). No-op when unlimited.
3919                    let cap = self.tile_cap.load(std::sync::atomic::Ordering::Relaxed);
3920                    if cap != u64::MAX {
3921                        g.set_cache_cap(cap);
3922                    }
3923                    Some(g)
3924                }
3925                Err(_) => {
3926                    self.fail();
3927                    None
3928                }
3929            }
3930        }
3931    }
3932
3933    /// Drop the decoded index of the graph named `iri`, if it has been opened,
3934    /// so its tiles and per-graph tile cache are freed. Takes `&mut self`:
3935    /// `OnceLock::take` needs unique access, which the export loop has between
3936    /// slots. Leaves the entry's `meta` (iri + range) so the graph re-opens
3937    /// lazily if touched again. A no-op if the graph is absent or never opened.
3938    fn release(&mut self, iri: &str) {
3939        let Some((_, table)) = self.dir.get_mut() else {
3940            return;
3941        };
3942        for slab in table.iter_mut() {
3943            let Some(entries) = slab.get_mut() else {
3944                continue;
3945            };
3946            for e in entries.iter_mut() {
3947                let matches = e
3948                    .meta
3949                    .get()
3950                    .map(|(name, _)| name.as_str() == iri)
3951                    .unwrap_or(false);
3952                if matches {
3953                    e.index.take();
3954                    return;
3955                }
3956            }
3957        }
3958    }
3959
3960    /// Visit every already-opened graph index (never triggers a fetch).
3961    fn for_each_opened(&self, mut f: impl FnMut(&GraphIndex)) {
3962        if let Some((_, table)) = self.dir.get() {
3963            for slab in table.iter().filter_map(|s| s.get()) {
3964                for e in slab.iter() {
3965                    if let Some(g) = e.index.get() {
3966                        f(g);
3967                    }
3968                }
3969            }
3970        }
3971    }
3972}
3973
3974/// The NAMED_GRAPHS section, held resident (in-memory and eager ranged opens —
3975/// unchanged) or walked lazily (the lazy ranged open, where fetching and
3976/// decoding every graph's index up front defeated remote laziness: 67 MB
3977/// fetched and ~32k indexes built before the first query on a many-graph file).
3978enum NamedGraphsSlot {
3979    Resident(Vec<(String, GraphIndex)>),
3980    Lazy(LazyNamedGraphs),
3981    /// The quad index (kind 10), on every open path: graphs are views over
3982    /// shared tiles, found through the dictionary and the sorted directory.
3983    Quad(crate::quad::QuadNamed),
3984}
3985
3986impl NamedGraphsSlot {
3987    /// Heap bytes the named-graph directory holds outside any cache: every
3988    /// walked entry's IRI and range, the walk buffer, and the directories of
3989    /// the graph indexes currently open (see [`GraphIndex::directory_bytes`]).
3990    fn directory_bytes(&self) -> u64 {
3991        match self {
3992            NamedGraphsSlot::Resident(v) => v
3993                .iter()
3994                .map(|(iri, g)| iri.capacity() as u64 + g.directory_bytes())
3995                .sum(),
3996            NamedGraphsSlot::Lazy(l) => {
3997                let mut bytes = l.walk.lock().map(|w| w.buf.capacity() as u64).unwrap_or(0);
3998                if let Some((_, table)) = l.dir.get() {
3999                    bytes += table.len() as u64
4000                        * std::mem::size_of::<std::sync::OnceLock<Box<[NamedEntry]>>>() as u64;
4001                    for slab in table.iter().filter_map(|s| s.get()) {
4002                        bytes += slab.len() as u64 * std::mem::size_of::<NamedEntry>() as u64;
4003                        for e in slab.iter() {
4004                            if let Some((iri, _)) = e.meta.get() {
4005                                bytes += iri.capacity() as u64;
4006                            }
4007                            if let Some(g) = e.index.get() {
4008                                bytes += g.directory_bytes();
4009                            }
4010                        }
4011                    }
4012                }
4013                bytes
4014            }
4015            NamedGraphsSlot::Quad(q) => q.directory_bytes(),
4016        }
4017    }
4018
4019    fn count(&self) -> usize {
4020        match self {
4021            NamedGraphsSlot::Resident(v) => v.len(),
4022            NamedGraphsSlot::Lazy(l) => l.count(),
4023            NamedGraphsSlot::Quad(q) => q.count(),
4024        }
4025    }
4026
4027    /// `exhaustive` is the walk-demand hint (resident opens ignore it): see
4028    /// [`LazyNamedGraphs::ensure_meta`]. `dict` names the graphs of a quad
4029    /// index (their names are dictionary terms there).
4030    fn name_at<'a>(&'a self, dict: &Dictionary, i: usize, exhaustive: bool) -> Option<&'a str> {
4031        match self {
4032            NamedGraphsSlot::Resident(v) => v.get(i).map(|(iri, _)| iri.as_str()),
4033            NamedGraphsSlot::Lazy(l) => l.name_at(i, exhaustive),
4034            NamedGraphsSlot::Quad(q) => q.name_at(dict, i, exhaustive),
4035        }
4036    }
4037
4038    /// [`name_at`](Self::name_at) without keeping the name: a quad-index walk
4039    /// over many graphs must not memoise one string per graph it passes.
4040    fn name_owned(&self, dict: &Dictionary, i: usize, exhaustive: bool) -> Option<String> {
4041        match self {
4042            NamedGraphsSlot::Quad(q) => q.name_owned(dict, i, exhaustive),
4043            _ => self.name_at(dict, i, exhaustive).map(str::to_owned),
4044        }
4045    }
4046
4047    fn graph_at<'a>(
4048        &'a self,
4049        dict: &Dictionary,
4050        i: usize,
4051        exhaustive: bool,
4052    ) -> Option<(&'a str, &'a GraphIndex)> {
4053        match self {
4054            NamedGraphsSlot::Resident(v) => v.get(i).map(|(iri, g)| (iri.as_str(), g)),
4055            NamedGraphsSlot::Lazy(l) => l.graph_at(i, exhaustive),
4056            NamedGraphsSlot::Quad(q) => q.graph_at(dict, i, exhaustive),
4057        }
4058    }
4059
4060    fn find<'a>(&'a self, dict: &Dictionary, iri: &str) -> Option<&'a GraphIndex> {
4061        match self {
4062            NamedGraphsSlot::Quad(q) => q.find(dict, iri),
4063            NamedGraphsSlot::Resident(v) => v.iter().find(|(name, _)| name == iri).map(|(_, g)| g),
4064            NamedGraphsSlot::Lazy(l) => {
4065                // Walk headers only until the IRI matches; decode just that
4066                // graph's index. An absent IRI costs a full header walk — the
4067                // layout has no by-name directory. A point lookup is targeted
4068                // demand: never hint the walk exhaustive.
4069                for i in 0..l.count() {
4070                    if l.name_at(i, false)? == iri {
4071                        return l.graph_at(i, false).map(|(_, g)| g);
4072                    }
4073                }
4074                None
4075            }
4076        }
4077    }
4078
4079    fn load_incomplete(&self) -> bool {
4080        match self {
4081            NamedGraphsSlot::Resident(v) => v.iter().any(|(_, g)| g.load_incomplete()),
4082            NamedGraphsSlot::Lazy(l) => {
4083                if l.failed.load(std::sync::atomic::Ordering::Relaxed) {
4084                    return true;
4085                }
4086                let mut bad = false;
4087                l.for_each_opened(|g| bad |= g.load_incomplete());
4088                bad
4089            }
4090            NamedGraphsSlot::Quad(q) => q.load_incomplete(),
4091        }
4092    }
4093
4094    fn reset_load_failures(&self) {
4095        match self {
4096            NamedGraphsSlot::Resident(v) => {
4097                for (_, g) in v {
4098                    g.reset_load_failure();
4099                }
4100            }
4101            NamedGraphsSlot::Lazy(l) => {
4102                l.failed.store(false, std::sync::atomic::Ordering::Relaxed);
4103                l.for_each_opened(|g| g.reset_load_failure());
4104            }
4105            NamedGraphsSlot::Quad(q) => q.reset_load_failures(),
4106        }
4107    }
4108
4109    /// The named-graph slot of a file opened with all its bytes in hand
4110    /// (`Rete::open`, `open_ranged`): the per-graph section decoded resident,
4111    /// or the quad index read through an owned copy of its section.
4112    fn resident(
4113        header: &Header,
4114        quad_section: impl FnOnce(u64, u64) -> Result<Vec<u8>, FileError>,
4115        per_graph: impl FnOnce() -> Result<Vec<(String, GraphIndex)>, FileError>,
4116    ) -> Result<Self, FileError> {
4117        check_named_layout(header)?;
4118        if let Some(q) = header.quad_index_section() {
4119            let bytes = quad_section(q.offset, q.length)?;
4120            return Ok(NamedGraphsSlot::Quad(crate::quad::QuadNamed::resident(
4121                bytes,
4122                header.block_codec,
4123                header.perms,
4124            )));
4125        }
4126        Ok(NamedGraphsSlot::Resident(if header.named_graphs_len > 0 {
4127            per_graph()?
4128        } else {
4129            Vec::new()
4130        }))
4131    }
4132}
4133
4134/// A read-only, in-memory view over a `.rete` file image.
4135pub struct Rete {
4136    header: Header,
4137    dict: Dictionary,
4138    index: GraphIndex,
4139    index_section_ranges: [ByteRange; NUM_PERMS],
4140    /// Per-permutation tile directories as absolute file ranges
4141    /// (`(min_a, max_a, compressed-tile range)`), for provenance. Empty for
4142    /// pre-tiling (v0.1) files.
4143    tile_ranges: [Vec<(u32, u32, ByteRange)>; NUM_PERMS],
4144    pyramid: PyramidSlot,
4145    text_index: TextIndexSlot,
4146    /// The LABEL_INDEX section (kind 8), if any.
4147    labels: SectionSlot<crate::label_index::LabelIndex>,
4148    /// The SCHEMA section (kind 9), if any.
4149    schema: SectionSlot<crate::schema_section::SchemaMeta>,
4150    named_graphs: NamedGraphsSlot,
4151    /// Raw bytes of the metadata section (empty if the file has none). The
4152    /// application layer decodes this (the CLI stores a JSON Dataset Card here).
4153    /// Only [`Rete::open`] populates it; [`Rete::open_ranged`] leaves it empty to
4154    /// preserve its minimal-fetch budget.
4155    metadata: Vec<u8>,
4156    /// Executes remote `SERVICE` blocks (SPARQL 1.1 federated query) — attached
4157    /// by the host via [`Rete::set_service_client`]; `None` means a non-SILENT
4158    /// `SERVICE` fails the query. Like the range readers, the engine never does
4159    /// I/O itself.
4160    service_client: Option<Box<dyn crate::service::ServiceClient>>,
4161    /// First failed non-SILENT `SERVICE` call of the current query. The row
4162    /// pipeline is infallible (the same contract as lazy tile fetches), so the
4163    /// failure is recorded here and taken by the top-level eval entry points.
4164    service_error: std::sync::Mutex<Option<String>>,
4165    /// Bytes a query's `COUNT(DISTINCT …)` set may hold before it spills to
4166    /// disk — see [`Rete::spill_threshold`].
4167    spill_threshold: std::sync::atomic::AtomicU64,
4168    /// The open `0x06` file (header, root, reader), for a `0x06` open.
4169    v06: Option<std::sync::Arc<crate::v06::open::V06File>>,
4170}
4171
4172impl Rete {
4173    /// Parse a full file image (v0 loads everything; a range-reading client
4174    /// will fetch only the sections it needs — same container format).
4175    pub fn open(bytes: &[u8]) -> Result<Self, FileError> {
4176        #[cfg(feature = "v06-parity")]
4177        if let Some(v6) = parity_image(bytes) {
4178            return Self::open(&v6);
4179        }
4180        if crate::v06::open::is_v06(bytes) {
4181            // A 0x06 image opens through the lazy machinery over the bytes in
4182            // memory: the same structures, faulted from RAM.
4183            let r: std::sync::Arc<dyn RangeReader + Send + Sync> =
4184                std::sync::Arc::new(crate::reader::BytesReader(std::sync::Arc::from(bytes)));
4185            let mut rete = Self::open_v06(r, bytes, false)?;
4186            rete.metadata = rete
4187                .v06
4188                .as_ref()
4189                .and_then(|f| f.card().ok().flatten())
4190                .unwrap_or_default();
4191            return Ok(rete);
4192        }
4193        let header = Header::from_bytes(bytes)?;
4194
4195        // Header offsets/lengths are untrusted (a `.rete` may be fetched truncated
4196        // or corrupt from an arbitrary URL). Slice through a checked helper so a
4197        // bad region yields an error instead of panicking on an OOB index.
4198        let region = |off: u64, len: u64| -> Result<&[u8], FileError> {
4199            checked_slice(bytes, off, len)
4200                .ok_or(FileError::Container("section range out of bounds"))
4201        };
4202
4203        let dict = decode_dictionary_container(
4204            region(header.dictionary_offset, header.dictionary_len)?,
4205            header.dict_codec,
4206        )?;
4207
4208        check_named_layout(&header)?;
4209        let index_bytes = region(header.root_dir_offset, header.root_dir_len)?;
4210        let (index, _) = decode_root_index_container(index_bytes, &header)?;
4211        let index_section_ranges = decode_index_section_ranges(
4212            index_bytes,
4213            header.root_dir_offset,
4214            header.perms,
4215            header.index_container_sections(),
4216        )?;
4217
4218        let pyramid = PyramidSlot::Resident(if header.pyramid_meta_len > 0 {
4219            Some(
4220                PyramidMeta::decode(region(header.pyramid_meta_offset, header.pyramid_meta_len)?)
4221                    .map_err(|_| FileError::Container("malformed pyramid meta"))?,
4222            )
4223        } else {
4224            None
4225        });
4226
4227        // The TEXT_INDEX section (opt-in `--text-index`); decode the whole thing
4228        // resident on a full-image open.
4229        let (ti_off, ti_len) = header.text_index_payload();
4230        let text_index = resident_text_index_slot(
4231            if ti_len > 0 {
4232                Some(region(ti_off, ti_len)?)
4233            } else {
4234                None
4235            },
4236            header.block_codec,
4237        )?;
4238        let labels = resident_label_slot(&header, &region);
4239        let schema = resident_schema_slot(&header, &region);
4240
4241        let named_graphs = NamedGraphsSlot::resident(
4242            &header,
4243            |off, len| Ok(region(off, len)?.to_vec()),
4244            || {
4245                decode_named_graphs(
4246                    region(header.named_graphs_offset, header.named_graphs_len)?,
4247                    header.block_codec,
4248                    header.perms,
4249                )
4250            },
4251        )?;
4252
4253        let metadata = if header.metadata_len > 0 {
4254            region(header.metadata_offset, header.metadata_len)?.to_vec()
4255        } else {
4256            Vec::new()
4257        };
4258
4259        let tile_ranges =
4260            tile_file_ranges(index_bytes, header.root_dir_offset, &index_section_ranges);
4261        Ok(Self {
4262            header,
4263            dict,
4264            index,
4265            index_section_ranges,
4266            tile_ranges,
4267            pyramid,
4268            text_index,
4269            labels,
4270            schema,
4271            named_graphs,
4272            metadata,
4273            service_client: None,
4274            service_error: std::sync::Mutex::new(None),
4275            spill_threshold: std::sync::atomic::AtomicU64::new(u64::MAX),
4276            v06: None,
4277        })
4278    }
4279
4280    /// Attach the client that executes `SERVICE <endpoint> { … }` blocks
4281    /// (SPARQL 1.1 federated query) against remote SPARQL endpoints. Without
4282    /// one, a non-SILENT `SERVICE` fails the query with a clear error and a
4283    /// `SERVICE SILENT` degrades to one empty solution, per the spec.
4284    pub fn set_service_client(&mut self, client: Box<dyn crate::service::ServiceClient>) {
4285        self.service_client = Some(client);
4286    }
4287
4288    pub(crate) fn service_client(&self) -> Option<&dyn crate::service::ServiceClient> {
4289        self.service_client.as_deref()
4290    }
4291
4292    /// Record a failed non-SILENT `SERVICE` call (first error wins).
4293    pub(crate) fn record_service_error(&self, msg: &str) {
4294        let mut e = self.service_error.lock().unwrap();
4295        if e.is_none() {
4296            *e = Some(msg.to_string());
4297        }
4298    }
4299
4300    /// Take (and clear) the pending `SERVICE` failure — called by every
4301    /// top-level eval entry so it can never leak into a later query.
4302    pub(crate) fn take_service_error(&self) -> Option<String> {
4303        self.service_error.lock().unwrap().take()
4304    }
4305
4306    pub fn header(&self) -> &Header {
4307        &self.header
4308    }
4309
4310    /// The file's byte layout, for visualization: header, metadata,
4311    /// dictionary, each index permutation's tile directory and individual
4312    /// tiles, pyramid summary, and named graphs — sorted by offset. Bytes not
4313    /// covered by any segment are container framing (section directories and
4314    /// length fields).
4315    pub fn file_layout(&self) -> Vec<LayoutSegment> {
4316        if let Some(f) = &self.v06 {
4317            return v06_layout(f);
4318        }
4319        let h = &self.header;
4320        let seg = |kind: &'static str, label: String, offset: u64, len: u64| LayoutSegment {
4321            kind,
4322            label,
4323            offset,
4324            len,
4325        };
4326        // The size in the label is DERIVED from the constant that defines the
4327        // span, never typed out: this line read "fixed 128 bytes" — the
4328        // pre-v0.3 header size — while emitting a 1024-byte span, and it is
4329        // what the file explorer shows a reader inspecting a real file.
4330        // `layout_header_label_matches_its_span` pins the two together.
4331        let mut out = vec![seg(
4332            "header",
4333            format!("header (fixed {} bytes)", crate::header::HEADER_LEN),
4334            0,
4335            crate::header::HEADER_LEN as u64,
4336        )];
4337        if h.metadata_len > 0 {
4338            out.push(seg(
4339                "metadata",
4340                "metadata (dataset card)".into(),
4341                h.metadata_offset,
4342                h.metadata_len,
4343            ));
4344        }
4345        out.push(seg(
4346            "dictionary",
4347            "dictionary (4 front-coded term sections)".into(),
4348            h.dictionary_offset,
4349            h.dictionary_len,
4350        ));
4351        for (si, perm) in crate::index::ALL_PERMS.into_iter().enumerate() {
4352            let sec = self.index_section_ranges[si];
4353            if sec.len == 0 {
4354                continue;
4355            }
4356            let first_tile = self.tile_ranges[si]
4357                .first()
4358                .map(|&(_, _, r)| r.offset)
4359                .unwrap_or(sec.offset + sec.len);
4360            if first_tile > sec.offset {
4361                out.push(seg(
4362                    "directory",
4363                    format!("{} tile directory", perm.name()),
4364                    sec.offset,
4365                    first_tile - sec.offset,
4366                ));
4367            }
4368            for (ti, &(min_a, max_a, r)) in self.tile_ranges[si].iter().enumerate() {
4369                out.push(seg(
4370                    "tile",
4371                    format!("{} tile {ti} (leading ids {min_a}..{max_a})", perm.name()),
4372                    r.offset,
4373                    r.len,
4374                ));
4375            }
4376        }
4377        if h.pyramid_meta_len > 0 {
4378            out.push(seg(
4379                "pyramid",
4380                "pyramid summary (communities + superedges)".into(),
4381                h.pyramid_meta_offset,
4382                h.pyramid_meta_len,
4383            ));
4384        }
4385        if h.named_graphs_len > 0 {
4386            out.push(seg(
4387                "named-graphs",
4388                format!("named graphs ({})", self.named_graphs.count()),
4389                h.named_graphs_offset,
4390                h.named_graphs_len,
4391            ));
4392        }
4393        if let Some(s) = h.label_index_section() {
4394            let label = match self.labels.get_if_loaded() {
4395                Some(li) => format!("label index ({} entries)", li.entry_count()),
4396                None => "label index".to_string(),
4397            };
4398            out.push(seg("label-index", label, s.offset, s.length));
4399        }
4400        if let Some(s) = h.schema_section() {
4401            out.push(seg(
4402                "schema",
4403                "schema (classes, query_stats, shapes)".into(),
4404                s.offset,
4405                s.length,
4406            ));
4407        }
4408        if let Some(q) = h.quad_index_section() {
4409            out.push(seg(
4410                "quad-index",
4411                format!("quad index ({} named graphs)", self.named_graphs.count()),
4412                q.offset,
4413                q.length,
4414            ));
4415        }
4416        out.sort_by_key(|s| s.offset);
4417        out
4418    }
4419
4420    /// Raw bytes of the file's metadata section, or `None` if it has none. The
4421    /// CLI stores a JSON Dataset Card here; `rete-core` treats it as opaque.
4422    /// Populated by [`Rete::open`] only — an [`Rete::open_ranged`] view returns
4423    /// `None` here (the card is not fetched on the minimal query path).
4424    pub fn metadata(&self) -> Option<&[u8]> {
4425        if self.metadata.is_empty() {
4426            None
4427        } else {
4428            Some(&self.metadata)
4429        }
4430    }
4431
4432    pub fn dictionary(&self) -> &Dictionary {
4433        &self.dict
4434    }
4435
4436    /// The pyramid metadata (summary graph + tiles), if the file has a pyramid.
4437    pub fn pyramid(&self) -> Option<&PyramidMeta> {
4438        match &self.pyramid {
4439            PyramidSlot::Resident(p) => p.as_ref(),
4440            // Faults the (possibly large) pyramid section on first access only.
4441            PyramidSlot::Lazy { loader, cell } => cell.get_or_init(loader).as_ref(),
4442        }
4443    }
4444
4445    /// The pyramid metadata **only if already resident or previously faulted** —
4446    /// never triggers a lazy range read. The query planner uses this for
4447    /// cardinality estimation so it is free for an in-memory file and never adds
4448    /// a fetch on the lazy remote path (which defers the pyramid by design).
4449    pub fn pyramid_if_loaded(&self) -> Option<&PyramidMeta> {
4450        match &self.pyramid {
4451            PyramidSlot::Resident(p) => p.as_ref(),
4452            PyramidSlot::Lazy { cell, .. } => cell.get().and_then(|o| o.as_ref()),
4453        }
4454    }
4455
4456    /// The SCHEMA section (kind 9): class hierarchy, level rollups and links,
4457    /// coherence axioms, `query_stats`, characteristic sets. Faults it in (one
4458    /// small range read) on the lazy path. `None` when the file has none — a
4459    /// file built before it existed keeps those fields in its pyramid-meta
4460    /// ([`schema_or_pyramid`](Self::schema_or_pyramid) reads either).
4461    pub fn schema(&self) -> Option<&crate::schema_section::SchemaMeta> {
4462        self.schema.get()
4463    }
4464
4465    /// The schema fields from the SCHEMA section when the file has one, else
4466    /// from the pyramid-meta (faulting either on the lazy path). `None` when
4467    /// the file carries neither.
4468    pub fn schema_or_pyramid(
4469        &self,
4470    ) -> Option<std::borrow::Cow<'_, crate::schema_section::SchemaMeta>> {
4471        if let Some(s) = self.schema() {
4472            return Some(std::borrow::Cow::Borrowed(s));
4473        }
4474        self.pyramid()
4475            .map(|p| std::borrow::Cow::Owned(crate::schema_section::SchemaMeta::from_pyramid(p)))
4476    }
4477
4478    /// Per-predicate planner statistics (`query_stats`): from the SCHEMA
4479    /// section (faulting it on the lazy path — it is small), else from the
4480    /// pyramid-meta when that is resident. Empty when the file has neither.
4481    /// See [`crate::meta::PredStat`].
4482    pub fn predicate_stats(&self) -> &[crate::meta::PredStat] {
4483        if let Some(s) = self.schema().filter(|s| !s.predicate_stats.is_empty()) {
4484            return &s.predicate_stats;
4485        }
4486        self.pyramid_if_loaded()
4487            .map(|p| p.predicate_stats.as_slice())
4488            .unwrap_or(&[])
4489    }
4490
4491    /// The statistics the planner may use **without a read it was not going to
4492    /// make**: the SCHEMA section if it is resident or already faulted — or,
4493    /// with `fault`, after faulting it (one small range, once per session) —
4494    /// else the resident pyramid-meta's block.
4495    pub(crate) fn planner_stats(&self, fault: bool) -> &[crate::meta::PredStat] {
4496        let schema = if fault {
4497            self.schema.get()
4498        } else {
4499            self.schema.get_if_loaded()
4500        };
4501        if let Some(s) = schema.filter(|s| !s.predicate_stats.is_empty()) {
4502            return &s.predicate_stats;
4503        }
4504        self.pyramid_if_loaded()
4505            .map(|p| p.predicate_stats.as_slice())
4506            .unwrap_or(&[])
4507    }
4508
4509    /// The entity shapes (characteristic sets): from the SCHEMA section, else
4510    /// from a resident pyramid-meta. See [`crate::meta::CharSet`].
4511    pub fn char_sets(&self) -> &[crate::meta::CharSet] {
4512        if let Some(s) = self.schema().filter(|s| !s.char_sets.is_empty()) {
4513            return &s.char_sets;
4514        }
4515        self.pyramid_if_loaded()
4516            .map(|p| p.char_sets.as_slice())
4517            .unwrap_or(&[])
4518    }
4519
4520    /// The **legacy** label block of a pyramid-meta (≤ 8,192 entries) — empty
4521    /// when the file has none or the pyramid isn't resident. Files built with
4522    /// `--label-index` carry a LABEL_INDEX section instead; see
4523    /// [`label_index_section`](Self::label_index_section). See
4524    /// [`crate::meta::LabelEntry`].
4525    pub fn label_index(&self) -> &[crate::meta::LabelEntry] {
4526        self.pyramid_if_loaded()
4527            .map(|p| p.label_index.as_slice())
4528            .unwrap_or(&[])
4529    }
4530
4531    /// Whether the file carries a LABEL_INDEX section (kind 8). Header only.
4532    pub fn has_label_index(&self) -> bool {
4533        self.header.label_index_section().is_some()
4534    }
4535
4536    /// The LABEL_INDEX section, faulting its head in on the lazy path.
4537    pub fn label_index_section(&self) -> Option<&crate::label_index::LabelIndex> {
4538        self.labels.get()
4539    }
4540
4541    /// Prefix-search the labels: the subjects whose label starts with `prefix`
4542    /// (case-insensitive), as `(label, subject term)`, capped at `limit`.
4543    /// Reads the LABEL_INDEX section when the file has one (every labelled
4544    /// subject, no dictionary read); otherwise the legacy pyramid-meta block,
4545    /// which **faults the pyramid** on the lazy path. Empty when the file
4546    /// carries neither.
4547    pub fn prefix_search(&self, prefix: &str, limit: usize) -> Vec<(String, String)> {
4548        if let Some(li) = self.labels.get() {
4549            return li.prefix_search(prefix, limit);
4550        }
4551        let Some(pyr) = self.pyramid() else {
4552            return Vec::new();
4553        };
4554        resolve_label_hits(&self.dict, pyr.prefix_search(prefix, limit))
4555    }
4556
4557    /// The full-text index (TEXT_INDEX section), faulting it in on first access
4558    /// on the lazy remote path. `None` when the file carries no text index.
4559    pub(crate) fn text_index(&self) -> Option<&crate::text_index::TextIndex> {
4560        self.text_index.index()
4561    }
4562
4563    /// Whether this file carries a full-text (TEXT_INDEX) section, i.e. it was
4564    /// built with `--text-index`. Cheap — reads the header, never faults.
4565    pub fn has_text_index(&self) -> bool {
4566        self.header.text_index_len > 0
4567    }
4568
4569    /// What the full-text index covers (graphs, predicates) — `None` when the
4570    /// file has none. Faults the token table on the lazy remote path, as a
4571    /// search does; the scope block lives in it.
4572    pub fn text_index_scope(&self) -> Option<crate::text_index::TextIndexScope> {
4573        self.text_index().map(|ti| ti.scope().clone())
4574    }
4575
4576    /// Byte length of the TEXT_INDEX section's leading **token table** — the
4577    /// length varint plus the compressed table, which is exactly what a first
4578    /// [`text_search`](Self::text_search) faults. `None` when the file carries
4579    /// no text index (or its head could not be read).
4580    ///
4581    /// This is the number to quote as the cost of a first search, NOT
4582    /// `header().text_index_len`: that counts the postings blob too, which is
4583    /// the bulk of the section and is only ever fetched one posting list at a
4584    /// time. On the published `epfl-infoscience.rete` the section is 195 MB and
4585    /// its token table 29 MB — a 6.5× difference.
4586    ///
4587    /// Costs nothing on a resident open (measured while the bytes were in hand);
4588    /// a lazy/ranged open pays one ≤10-byte range read, memoized, and still
4589    /// never faults the table itself.
4590    pub fn text_index_token_table_len(&self) -> Option<u64> {
4591        self.text_index.token_table_len()
4592    }
4593
4594    /// Full-text search over the literals: subject IRIs that carry **every** word
4595    /// in `words` (whole-word, case-insensitive — AND semantics), optionally also
4596    /// requiring a word that **starts with** `prefix` (token-prefix). Results are
4597    /// ordered by subject id and capped at `limit` (0 = uncapped). Empty when the
4598    /// file has no text index or nothing matches.
4599    ///
4600    /// Like [`prefix_search`](Self::prefix_search), this **faults** the index on
4601    /// the lazy remote path — a search is an explicit read, and only the queried
4602    /// posting lists are fetched, not the whole index.
4603    pub fn text_search(&self, words: &[&str], prefix: Option<&str>, limit: usize) -> Vec<String> {
4604        text_search_in(self.text_index(), &self.dict, words, prefix, limit)
4605    }
4606
4607    /// The default-graph permutation index.
4608    pub fn default_index(&self) -> &GraphIndex {
4609        &self.index
4610    }
4611
4612    /// Resolve every triple of a graph (`None` = default graph) back to terms.
4613    pub fn dump(&self, graph: Option<&str>) -> Vec<TermTriple> {
4614        let index = match graph {
4615            None => &self.index,
4616            Some(g) => match self.graph_index(g) {
4617                Some(i) => i,
4618                None => return Vec::new(),
4619            },
4620        };
4621        index
4622            .match_pattern((None, None, None))
4623            .into_iter()
4624            .filter_map(|(s, p, o)| {
4625                Some((
4626                    self.dict.subject_term(s)?,
4627                    self.dict.predicate_term(p)?,
4628                    self.dict.object_term(o)?,
4629                ))
4630            })
4631            .collect()
4632    }
4633
4634    /// Stream every triple of a graph (`None` = default) to `f`, resolving terms
4635    /// a batch at a time — no full `Vec` materialization, so it is safe on graphs far
4636    /// larger than RAM. `rete export` uses this to serialize 100M+ triple files
4637    /// that `dump()` (which collects every term into a `Vec<String>`) would OOM on.
4638    ///
4639    /// On a lazy/ranged open the dictionary is **not** prefetched whole: each
4640    /// batch faults only the chunks its own terms live in. Dumping ONE named
4641    /// graph therefore costs that graph's terms, not the file's. Dumping
4642    /// *everything* still reads every dictionary byte and leaves every chunk
4643    /// resident — that is inherent, and it, not the prefetch, is what sets the
4644    /// peak of a full export. See the comment in the body for both measurements.
4645    ///
4646    /// To dump a *slice* — one predicate, one subject, one object — call
4647    /// [`dump_filtered_each`](Self::dump_filtered_each), which is this method
4648    /// with a triple pattern and prunes tiles instead of filtering rows.
4649    pub fn dump_each<F: FnMut(&str, &str, &str)>(&self, graph: Option<&str>, f: F) {
4650        self.dump_filtered_each(graph, None, None, None, f)
4651    }
4652
4653    /// [`dump_each`](Self::dump_each) restricted to a triple **pattern** — the
4654    /// filtered dump. Any of `s` / `p` / `o` may be bound (canonical N-Triples
4655    /// term tokens); all three unbound is exactly `dump_each`.
4656    ///
4657    /// # Why this is cheap, and how much
4658    ///
4659    /// It is the same access path a routed *query* takes, not a full scan with a
4660    /// predicate test bolted on: [`GraphIndex::scan_iter`] picks the permutation
4661    /// with the longest bound prefix, binary-searches that section's tile
4662    /// directory down to the span a bound leading id can live in, and then drops
4663    /// every remaining tile whose recorded synopsis proves it cannot hold a
4664    /// bound secondary component — **all from directories, without fetching a
4665    /// tile**. So a filtered dump pays for the slice, which is the whole point
4666    /// of issue #117.
4667    ///
4668    /// Measured on `cordis.rete` (801 MB, a 417 MB dictionary, 26.4M quads,
4669    /// six named graphs), lazily opened; "before" is the only way to get a
4670    /// predicate slice out of a dump previously — dump the graph and throw away
4671    /// what does not match, in the consumer:
4672    ///
4673    /// ```text
4674    ///   graph=results  p=s66#doi     -> 337,811 rows
4675    ///     before  375.8 MB · 1797 req · 2105.8 MB RSS · 12.8 s · first row 409 ms
4676    ///     after    16.0 MB ·  105 req ·  182.6 MB RSS ·  0.41 s · first row  14 ms
4677    ///   graph=results  p=s66#isbn    ->  40,761 rows
4678    ///     before  375.8 MB · 1797 req · 2105.7 MB RSS · 12.8 s
4679    ///     after    15.5 MB ·  104 req ·  155.2 MB RSS ·  0.23 s
4680    /// ```
4681    ///
4682    /// # Where it does NOT help
4683    ///
4684    /// The index is pruned; the **dictionary is not**. Term resolution still
4685    /// faults whichever chunks the surviving rows' terms live in, and on a file
4686    /// whose payload *is* the literals those chunks are most of the file. Same
4687    /// file, a predicate whose objects are the long abstracts:
4688    ///
4689    /// ```text
4690    ///   graph=projects p=s66#abstract ->  80,206 rows
4691    ///     before  260.7 MB · 1142 req · 1496.9 MB RSS · 5.97 s
4692    ///     after   213.4 MB · 1606 req · 1144.4 MB RSS · 2.91 s   (1.2x, not 23x)
4693    /// ```
4694    ///
4695    /// An *unfiltered* dump is unchanged by construction — it is the floor, and
4696    /// its peak is the resident dictionary, not this. Two ceilings sit above
4697    /// both: a faulted dictionary chunk is a `OnceCell` nothing ever evicts, and
4698    /// on a literal-heavy file the chunk directory alone can be a third of the
4699    /// file before any dump work at all (#198).
4700    ///
4701    /// # Permutations
4702    ///
4703    /// Every one of the eight bound/unbound shapes routes inside
4704    /// [`PermSet::CORE`] — `{SPO, POS, OSP}`, the minimum a legal file carries
4705    /// (see `perm_routing_never_leaves_core`). A file built with
4706    /// `--permutations 3` therefore prunes *identically* to a six-permutation
4707    /// one: same tiles, same rows, no fallback path to get wrong.
4708    ///
4709    /// A bound term the dictionary does not know, or an absent graph IRI, yields
4710    /// nothing without touching the index.
4711    ///
4712    /// # Order
4713    ///
4714    /// Rows arrive in the **routed permutation's** order, which is the price of
4715    /// routing at all. Unfiltered (and subject-bound) that is SPO — canonical
4716    /// `(s, p, o)`, unchanged from [`dump_each`](Self::dump_each). A bound
4717    /// predicate routes to POS and streams `(p, o, s)`; a bound object routes to
4718    /// OSP. The *set* is identical to the unfiltered dump filtered; the order is
4719    /// not, so a consumer that needs canonical order must sort. N-Quads, the
4720    /// format `rete export` writes, is order-independent.
4721    ///
4722    /// [`PermSet::CORE`]: crate::index::PermSet::CORE
4723    pub fn dump_filtered_each<F: FnMut(&str, &str, &str)>(
4724        &self,
4725        graph: Option<&str>,
4726        s: Option<&str>,
4727        p: Option<&str>,
4728        o: Option<&str>,
4729        mut f: F,
4730    ) {
4731        let index = match graph {
4732            None => Some(&self.index),
4733            Some(g) => self.graph_index(g),
4734        };
4735        let Some(index) = index else { return };
4736        // A bound term the dictionary does not know can match nothing, and the
4737        // index is never touched — same contract as `query_iter`/`query_batch`.
4738        let Some(pattern) = self.resolve_query_pattern(s, p, o) else {
4739            return;
4740        };
4741        let dict = &self.dict;
4742        // Resolve in fixed-size batches with ONE coalesced dictionary fault per
4743        // batch — the same `prefetch_terms` call `dump_batch` makes, for the
4744        // same reason.
4745        //
4746        // This used to be `Dictionary::prefetch_all` up front, on the argument
4747        // that a full dump reaches every term anyway so the bytes are owed
4748        // either way. That argument only holds for a dump of EVERYTHING, and
4749        // this takes a graph AND a pattern. Measured on `cordis.rete` (801 MB,
4750        // a 417 MB dictionary, six named graphs), lazily opened:
4751        //
4752        //   dump_each(Some(fundingschemes))  417 MB read, 2510 MB peak RSS
4753        //                              →     173 MB read,  619 MB peak RSS
4754        //   dump_each(None) — the EMPTY default graph, which is the first thing
4755        //   `rete export` calls on a quads file:
4756        //                                    415 MB read, 2508 MB peak RSS
4757        //                              →       0 MB read,   95 MB peak RSS
4758        //
4759        // The 2.5 GB comes from asking for every chunk at once: the bulk loader
4760        // coalesces adjacent chunks into one span, so the whole dictionary
4761        // section is materialized three times over before the first triple —
4762        // the fetched span, the per-chunk copies sliced out of it, and the
4763        // decompressed bodies.
4764        //
4765        // Be clear about what this does NOT buy. A dump of the whole default
4766        // graph still faults every chunk, and chunks stay resident once
4767        // faulted, so `rete export` of a single-graph file is unchanged:
4768        // `figshare.rete` (222 MB, a 126 MB dictionary) peaks at 874 MB lazily
4769        // either way. The resident dictionary is the floor there, not this.
4770        //
4771        // Within each window, term resolution is CHUNK-ORDERED, not row-ordered
4772        // (`Dictionary::resolve_window`, bounded-export phase 1): the window's
4773        // ids are grouped by the `(section, chunk)` they live in and each chunk
4774        // is decoded once and its runs walked once, filling every slot that
4775        // chunk owns, then the rows are emitted in the original scan order. Two
4776        // things follow. First, a window's objects — which scatter across the
4777        // object-only section for one subject — touch each chunk a single time
4778        // instead of re-faulting it per row. Second, the per-id re-walk from a
4779        // run's start that the old row-order path did for *every* id collapses
4780        // to one monotonic forward pass per run. The output is byte-for-byte
4781        // identical (same bytes per id, same emission order, same
4782        // all-three-resolved emit rule); the cap is still unlimited, so this is
4783        // structural preparation for the real bound (phase 2), plus a small
4784        // free speedup. The coalesced bulk fault (`prefetch_terms`) is
4785        // unchanged, so the read pattern and peak residency are as before.
4786        const WINDOW: usize = 4096;
4787        // The whole-window coalesced prefetch (`prefetch_terms`) materializes
4788        // every distinct chunk the window touches at once — up to ~2·WINDOW
4789        // chunks for scattered objects. Under a GENEROUS budget that transient
4790        // fits and the coalesced reads are the fast path (small/normal files,
4791        // `--in-memory`, unlimited). Under a TIGHT budget it would blow the cap
4792        // and thrash (insert-then-evict every chunk, then re-fault it in
4793        // `resolve_window`), so skip it: `resolve_window` is chunk-ordered and
4794        // holds one chunk `Arc` at a time (plan §3.5), keeping the window's
4795        // working set to a single chunk regardless of how the objects scatter.
4796        // Output is byte-identical either way — prefetch only warms the cache.
4797        let cap = dict.cache_cap();
4798        let prefetch_window =
4799            cap == u64::MAX || cap >= (WINDOW as u64) * 2 * (DICT_CHUNK_BUDGET as u64);
4800        // Under a finite budget the prefetch also goes in batches of
4801        // PREFETCH_BATCH chunks: one bulk read holds everything it fetched
4802        // (span, slices, decoded bodies) until it is all cached, which for a
4803        // scattered window was thousands of chunks — up to ~1 GB on
4804        // openaire-relation — outside the cache cap. Unlimited keeps one read
4805        // per section (the latency-optimal shape for a remote file).
4806        const PREFETCH_BATCH: usize = 64;
4807        let prefetch_batch = if cap == u64::MAX {
4808            usize::MAX
4809        } else {
4810            PREFETCH_BATCH
4811        };
4812        let mut ids: Vec<(u32, u32, u32)> = Vec::with_capacity(WINDOW);
4813        let mut nodes: Vec<u32> = Vec::with_capacity(WINDOW * 2);
4814        let mut preds: Vec<u32> = Vec::with_capacity(WINDOW);
4815        let mut resolver = crate::dictionary::WindowResolver::default();
4816        let mut flush = |ids: &mut Vec<(u32, u32, u32)>,
4817                         resolver: &mut crate::dictionary::WindowResolver,
4818                         f: &mut F| {
4819            if ids.is_empty() {
4820                return;
4821            }
4822            // Coalesced bulk fault of exactly the chunks this window needs —
4823            // only under a generous budget (see `prefetch_window` above); a
4824            // tight budget faults chunk-by-chunk inside `resolve_window`.
4825            if prefetch_window {
4826                nodes.clear();
4827                preds.clear();
4828                for &(s, p, o) in ids.iter() {
4829                    nodes.push(dict.subject_node(s));
4830                    nodes.push(dict.object_node(o));
4831                    preds.push(p);
4832                }
4833                dict.prefetch_terms_batched(&nodes, &preds, prefetch_batch);
4834            }
4835            dict.resolve_window(ids, resolver, |s, p, o| f(s, p, o));
4836            ids.clear();
4837        };
4838        for t in index.scan_iter(pattern) {
4839            ids.push(t);
4840            if ids.len() == WINDOW {
4841                flush(&mut ids, &mut resolver, &mut f);
4842            }
4843        }
4844        flush(&mut ids, &mut resolver, &mut f);
4845    }
4846
4847    /// What a [`dump_filtered_each`](Self::dump_filtered_each) over this graph
4848    /// and pattern will fetch, **before** it starts — the dump twin of
4849    /// `rete cost`'s query preview.
4850    ///
4851    /// Costed from the tile directories, so it fetches no tile. It does read the
4852    /// dictionary far enough to resolve the bound terms (that is a real,
4853    /// unavoidable cost of the dump too) and to learn whether the graph exists.
4854    pub fn dump_plan(
4855        &self,
4856        graph: Option<&str>,
4857        s: Option<&str>,
4858        p: Option<&str>,
4859        o: Option<&str>,
4860    ) -> DumpPlan {
4861        let index = match graph {
4862            None => Some(&self.index),
4863            Some(g) => self.graph_index(g),
4864        };
4865        let scan = index
4866            .zip(self.resolve_query_pattern(s, p, o))
4867            .map(|(ix, pattern)| ix.scan_plan(pattern));
4868        DumpPlan {
4869            scan,
4870            dictionary_bytes: match &self.v06 {
4871                Some(f) => f.section(crate::v06::kind::DICT_DATA).map_or(0, |r| r.len),
4872                None => self.header.dictionary_len,
4873            },
4874        }
4875    }
4876
4877    /// The **pull** twin of [`Rete::dump_each`](Self::dump_each): every triple of a
4878    /// graph (`None` = default) as a lazy iterator of resolved terms.
4879    ///
4880    /// Same constant-memory scan — one triple decoded and resolved per `next()`,
4881    /// never a `Vec` of the whole graph — but the caller drives it, so it can
4882    /// stop early or be **suspended and resumed** across a foreign-function
4883    /// boundary. That is what the wasm/JS client's batched quad cursor needs:
4884    /// a callback cannot be paused mid-scan to hand control back to JavaScript,
4885    /// an iterator can.
4886    ///
4887    /// The dictionary is NOT prefetched: terms fault in as they are resolved, so
4888    /// stopping after a handful of triples costs a handful of dictionary reads
4889    /// rather than the whole dictionary — which on a lazy remote open of a large
4890    /// file would be gigabytes before the first triple, defeating the point of a
4891    /// pull API. [`dump_each`](Self::dump_each) does not prefetch it whole
4892    /// either; it batches the same per-term faults. Peak memory is
4893    /// O(faulted dictionary chunks + faulted index tiles), *not* O(triples).
4894    pub fn dump_iter(&self, graph: Option<&str>) -> impl Iterator<Item = TermTriple> + '_ {
4895        self.query_iter(graph, None, None, None)
4896    }
4897
4898    /// [`dump_iter`](Self::dump_iter) generalized to a triple **pattern**: the
4899    /// lazy pull form of [`query_in_graph`](Self::query_in_graph). A bound term
4900    /// unknown to the dictionary, or an unknown graph IRI, yields nothing.
4901    ///
4902    /// Same constant-memory guarantee — one triple decoded and resolved per
4903    /// `next()`, never a `Vec` of every match — and the same absence of a
4904    /// dictionary prefetch, so stopping after a handful of matches costs a
4905    /// handful of dictionary reads. `query_in_graph` is the eager twin; prefer
4906    /// this one wherever the consumer can stop early (`ASK`, `LIMIT`, an RDF4J
4907    /// `getStatements` that is about to be sliced).
4908    pub fn query_iter(
4909        &self,
4910        graph: Option<&str>,
4911        s: Option<&str>,
4912        p: Option<&str>,
4913        o: Option<&str>,
4914    ) -> impl Iterator<Item = TermTriple> + '_ {
4915        let index = match graph {
4916            None => Some(&self.index),
4917            Some(g) => self.graph_index(g),
4918        };
4919        index
4920            .zip(self.resolve_query_pattern(s, p, o))
4921            .into_iter()
4922            .flat_map(|(ix, pattern)| ix.scan_iter(pattern))
4923            .filter_map(move |(s, p, o)| {
4924                Some((
4925                    self.dict.subject_term(s)?,
4926                    self.dict.predicate_term(p)?,
4927                    self.dict.object_term(o)?,
4928                ))
4929            })
4930    }
4931
4932    /// One bounded, **resumable** slice of a pattern's matches inside one graph
4933    /// (`None` = default): `(triples, next_cursor, done)`. Start at `cursor = 0`
4934    /// and feed the returned cursor back until `done`.
4935    ///
4936    /// This is [`query_in_graph`](Self::query_in_graph) made streamable across a
4937    /// boundary that cannot hold a Rust borrow — the wasm/JVM one. The engine
4938    /// keeps no iterator alive between calls; the entire resume state is the
4939    /// opaque `u64` (see [`GraphIndex::scan_batch`]), so a client's cursor
4940    /// survives being suspended, and a cursor that is abandoned leaks nothing
4941    /// but a `u64`.
4942    ///
4943    /// The dictionary is faulted **per batch**, not whole: taking ten matches
4944    /// off the front of a 9.8-billion-quad file reads ten matches' worth of
4945    /// terms. That is the difference between `LIMIT 1` answering in bounded time
4946    /// and it reading a 30 GB dictionary first.
4947    ///
4948    /// `max_quads` is a floor (batches end on an a-group boundary), and every
4949    /// call either returns at least one row or reports `done`.
4950    pub fn query_batch(
4951        &self,
4952        graph: Option<&str>,
4953        s: Option<&str>,
4954        p: Option<&str>,
4955        o: Option<&str>,
4956        cursor: u64,
4957        max_quads: usize,
4958    ) -> (Vec<TermTriple>, u64, bool) {
4959        let index = match graph {
4960            None => &self.index,
4961            // An unknown graph IRI is an empty scan, not an error — same as
4962            // `query_in_graph` and `dump_batch`.
4963            Some(g) => match self.graph_index(g) {
4964                Some(i) => i,
4965                None => return (Vec::new(), cursor, true),
4966            },
4967        };
4968        // A bound term the dictionary does not know can match nothing, and the
4969        // index is never touched.
4970        let Some(pattern) = self.resolve_query_pattern(s, p, o) else {
4971            return (Vec::new(), cursor, true);
4972        };
4973        let (ids, next, done) = index.scan_batch(pattern, cursor, max_quads);
4974        // One coalesced dictionary fault for THIS batch — the same call
4975        // `query_in_graph` and `dump_batch` make, and the reason a partial walk
4976        // costs in proportion to what it actually read.
4977        if !ids.is_empty() {
4978            let mut nodes = Vec::with_capacity(ids.len() * 2);
4979            let mut preds = Vec::with_capacity(ids.len());
4980            for &(s, p, o) in &ids {
4981                nodes.push(self.dict.subject_node(s));
4982                nodes.push(self.dict.object_node(o));
4983                preds.push(p);
4984            }
4985            self.dict.prefetch_terms(&nodes, &preds);
4986        }
4987        let triples = ids
4988            .into_iter()
4989            .filter_map(|(s, p, o)| {
4990                Some((
4991                    self.dict.subject_term(s)?,
4992                    self.dict.predicate_term(p)?,
4993                    self.dict.object_term(o)?,
4994                ))
4995            })
4996            .collect();
4997        (triples, next, done)
4998    }
4999
5000    /// Resolve one bounded slice of a graph's triples. Returns
5001    /// `(triples, next_cursor, done)`; start with `cursor = 0` and feed the
5002    /// returned cursor back until `done`.
5003    ///
5004    /// This is the PULL half of the dump, and it exists because [`Rete::dump_each`] is
5005    /// push-based: it drives the scan itself and hands each triple to a
5006    /// callback. A client that wants to pull (a Python generator, a JS async
5007    /// iterator) has to hold the scan's stack between calls — which means a
5008    /// thread (unavailable on Pyodide and in wasm) or a self-referential struct
5009    /// holding an iterator that borrows the `Rete` stored next to it, whose
5010    /// soundness rests on drop order plus an unstated promise that this crate's
5011    /// lazily-faulted caches stay write-once. Neither is worth it for a dump.
5012    ///
5013    /// Instead the scan is made RESUMABLE: SPO tiles are ordered by subject id,
5014    /// so "every triple of subject `sid`, for ascending `sid`" visits exactly the
5015    /// tiles a full scan visits, in the same order, and the whole resume state
5016    /// collapses to one `u32`. No thread, no `unsafe`, no borrow held across
5017    /// calls — and unlike `(offset, limit)` it is O(n) overall rather than
5018    /// O(n²/limit), because nothing is ever re-scanned.
5019    ///
5020    /// `max_quads` is the FLOOR of a batch, not a hard cut: a batch always ends
5021    /// on a subject boundary, so no subject is split across two calls. Work per
5022    /// call is bounded by a probe budget too — a sparse named graph whose few
5023    /// subjects are spread over a huge id space returns early with `done = false`
5024    /// rather than grinding through absent ids, and the caller just asks again.
5025    pub fn dump_batch(
5026        &self,
5027        graph: Option<&str>,
5028        cursor: u32,
5029        max_quads: usize,
5030    ) -> (Vec<TermTriple>, u32, bool) {
5031        /// The first subject id at or after `sid` that any SPO tile can hold, or
5032        /// `None` when `sid` is past the last tile. Tiles ascend by leading
5033        /// range, so this is a binary search; it returns `sid` itself when a tile
5034        /// already covers it (an ordinary "this subject has no triples" step).
5035        fn next_tile_start(tiles: &crate::index::TileList, sid: u32) -> Option<u32> {
5036            let i = tiles.first_max_ge(sid);
5037            (i < tiles.len()).then(|| tiles[i].leading_range().0.max(sid))
5038        }
5039
5040        let index = match graph {
5041            None => self.default_index(),
5042            // An unknown graph IRI is an empty dump, not an error — same as
5043            // `dump_each`.
5044            Some(g) => match self.graph_index(g) {
5045                Some(i) => i,
5046                None => return (Vec::new(), cursor, true),
5047            },
5048        };
5049        // Tile leading ranges live in the section directory, so both the span and
5050        // the gap jumps below are free on a lazy/remote open — no tile is fetched
5051        // to decide where to look.
5052        let tiles = index.tile_sections()[IndexPermutation::Spo.section_index()];
5053        let (Some(first), Some(last)) = (
5054            tiles.first().map(|t| t.leading_range().0),
5055            tiles.last().map(|t| t.leading_range().1),
5056        ) else {
5057            return (Vec::new(), cursor, true); // empty graph
5058        };
5059
5060        let mut sid = cursor.max(first);
5061        let mut probes = max_quads.saturating_mul(4).max(1 << 16);
5062        let mut ids: Vec<(u32, u32, u32)> = Vec::new();
5063        let mut exhausted = false;
5064        while sid <= last && probes > 0 {
5065            probes -= 1;
5066            let hits = index.match_pattern((Some(sid), None, None));
5067            if hits.is_empty() {
5068                // No triples at this id: jump to the next subject the index
5069                // holds (the tile's group directory knows it), instead of
5070                // stepping through absent ids one at a time. That is what keeps
5071                // a quad-index view — whose tile shared with other graphs has
5072                // the conservative range `0..=u32::MAX` — from probing four
5073                // billion ids. Only when that lookup cannot answer (a failed
5074                // fetch, already recorded) does it fall back to the directory:
5075                // if the id falls in a hole BETWEEN tiles, jump to the next
5076                // tile's first id.
5077                match index.next_leading(IndexPermutation::Spo.section_index(), sid) {
5078                    Some(next) if next > sid => {
5079                        sid = next;
5080                        continue;
5081                    }
5082                    None if !index.load_incomplete() => {
5083                        exhausted = true;
5084                        break;
5085                    }
5086                    _ => {}
5087                }
5088                match next_tile_start(tiles, sid) {
5089                    None => {
5090                        exhausted = true;
5091                        break;
5092                    }
5093                    Some(next) if next > sid => {
5094                        sid = next;
5095                        continue;
5096                    }
5097                    Some(_) => {}
5098                }
5099            } else {
5100                ids.extend(hits);
5101            }
5102            if sid == u32::MAX {
5103                exhausted = true;
5104                break;
5105            }
5106            sid += 1;
5107            if ids.len() >= max_quads {
5108                break;
5109            }
5110        }
5111        let done = exhausted || sid > last;
5112
5113        // One coalesced dictionary fault for this batch, then resolve.
5114        // Deliberately NOT `Dictionary::prefetch_all` (what `dump_each` uses):
5115        // on a lazy open that pulls the ENTIRE dictionary — gigabytes on a big
5116        // file — before the first quad, so taking five quads off the front would
5117        // cost as much as a full dump. Per-batch prefetch keeps a partial walk
5118        // proportional to what it actually read.
5119        let dict = self.dictionary();
5120        if !ids.is_empty() {
5121            let mut nodes = Vec::with_capacity(ids.len() * 2);
5122            let mut preds = Vec::with_capacity(ids.len());
5123            for &(s, p, o) in &ids {
5124                nodes.push(dict.subject_node(s));
5125                nodes.push(dict.object_node(o));
5126                preds.push(p);
5127            }
5128            dict.prefetch_terms(&nodes, &preds);
5129        }
5130        let triples = ids
5131            .into_iter()
5132            .filter_map(|(s, p, o)| {
5133                Some((
5134                    dict.subject_term(s)?,
5135                    dict.predicate_term(p)?,
5136                    dict.object_term(o)?,
5137                ))
5138            })
5139            .collect();
5140        (triples, sid, done)
5141    }
5142
5143    /// How many named graphs this dataset has. On a lazy ranged open this
5144    /// reads only the section's leading count varint — never the graphs.
5145    pub fn named_graph_count(&self) -> usize {
5146        self.named_graphs.count()
5147    }
5148
5149    /// The `i`-th named graph's quad count, when the file records it: a
5150    /// quad-index file keeps it in its graph directory, so this reads one
5151    /// directory block and never the graph's index. `None` for a per-graph
5152    /// (NAMED_GRAPHS section) file, or past the last graph.
5153    pub fn named_graph_quad_count(&self, i: usize) -> Option<u64> {
5154        match &self.named_graphs {
5155            NamedGraphsSlot::Quad(q) => q.quad_count_at(i),
5156            _ => None,
5157        }
5158    }
5159
5160    /// The `i`-th named graph as an **owned** index that is not memoised,
5161    /// when the file can make one cheaply: a quad-index file's graph of at
5162    /// most [`TRANSIENT_GRAPH_QUADS`] quads. A walk over every graph (`GRAPH
5163    /// ?g`, the union default graph, `query_quads`) evaluates each such graph
5164    /// and drops it, instead of keeping one index per graph for the life of
5165    /// the open — 1.5 M of them on the OpenCitations sample. `None` for a
5166    /// larger graph, a per-graph file, or a failed read: use
5167    /// [`named_graph_at`](Self::named_graph_at) then.
5168    pub fn named_graph_transient(&self, i: usize) -> Option<GraphIndex> {
5169        match &self.named_graphs {
5170            NamedGraphsSlot::Quad(q) => q.transient_view(i, TRANSIENT_GRAPH_QUADS),
5171            _ => None,
5172        }
5173    }
5174
5175    /// The ordinals (ascending, as [`named_graph_at`](Self::named_graph_at)
5176    /// counts them) of the named graphs holding at least one quad whose
5177    /// subject is one of `subjects` (subject ids), for a quad-index file: one
5178    /// sequential pass over its graph-leading SPO tiles, never a view per
5179    /// graph. `None` for a per-graph file, which has no such pass, or on a
5180    /// failed read (recorded, see [`index_incomplete`](Self::index_incomplete)).
5181    pub(crate) fn named_graphs_with_subjects(
5182        &self,
5183        subjects: &std::collections::BTreeSet<u32>,
5184    ) -> Option<Vec<usize>> {
5185        match &self.named_graphs {
5186            NamedGraphsSlot::Quad(q) => q.graphs_with_subjects(subjects),
5187            _ => None,
5188        }
5189    }
5190
5191    /// Deduplicated quads across every named graph, when the file records it
5192    /// (a quad-index file, from its preamble: one small read). `None` for a
5193    /// per-graph file — there it is the header's `quad_count` minus the
5194    /// default graph's triples.
5195    pub fn named_quad_count(&self) -> Option<u64> {
5196        match &self.named_graphs {
5197            NamedGraphsSlot::Quad(q) => Some(q.quad_count()),
5198            _ => None,
5199        }
5200    }
5201
5202    /// How many triples `index` holds, for sizing a pattern's tile estimate:
5203    /// a named graph's own count when it is known without a tile read (a
5204    /// quad-index graph's directory count, a resident per-graph index), else
5205    /// the header's total. The default graph keeps the total: its size is not
5206    /// recorded in the header, and the planner's thresholds were tuned on it.
5207    pub(crate) fn index_triples(&self, index: &GraphIndex) -> u64 {
5208        index.size_hint().unwrap_or(self.header().quad_count)
5209    }
5210
5211    /// The `i`-th named graph's IRI (stored order). On a lazy ranged open this
5212    /// walks entry HEADERS up to `i` — it never decodes a graph's index.
5213    pub fn named_graph_name_at(&self, i: usize) -> Option<&str> {
5214        self.named_graphs.name_at(&self.dict, i, false)
5215    }
5216
5217    /// The `i`-th named graph as `(iri, index)`. On a lazy ranged open the
5218    /// index is fetched and decoded on first access and memoised; check
5219    /// [`index_incomplete`](Self::index_incomplete) after evaluating.
5220    pub fn named_graph_at(&self, i: usize) -> Option<(&str, &GraphIndex)> {
5221        self.named_graphs.graph_at(&self.dict, i, false)
5222    }
5223
5224    /// The `i`-th graph's name as an owned string that the file does **not**
5225    /// keep: what a walk over millions of graphs uses, so the names it has
5226    /// passed are not all held at once. `exhaustive = true` tells a lazy walk
5227    /// the caller will visit every graph, so it may read the section in bulk
5228    /// chunks; purely a fetch strategy, results are identical either way. Only
5229    /// the SPARQL evaluator sets it, from shapes where full consumption is
5230    /// provable.
5231    pub(crate) fn named_graph_name_owned_demand(
5232        &self,
5233        i: usize,
5234        exhaustive: bool,
5235    ) -> Option<String> {
5236        self.named_graphs.name_owned(&self.dict, i, exhaustive)
5237    }
5238
5239    /// The ordinal of the quad-index graph named `iri`: a dictionary lookup and
5240    /// a directory binary search. `None` for another file layout, or no such
5241    /// graph.
5242    pub(crate) fn named_graph_ordinal(&self, iri: &str) -> Option<usize> {
5243        match &self.named_graphs {
5244            NamedGraphsSlot::Quad(q) => q.ordinal_of_iri(&self.dict, iri),
5245            _ => None,
5246        }
5247    }
5248
5249    /// The ordinal of the quad-index graph whose subject id is `g` (as
5250    /// [`named_quad_scan`](Self::named_quad_scan) yields it).
5251    pub(crate) fn named_graph_ordinal_of_id(&self, g: u32) -> Option<usize> {
5252        match &self.named_graphs {
5253            NamedGraphsSlot::Quad(q) => q.ordinal_of_id(g),
5254            _ => None,
5255        }
5256    }
5257
5258    /// A streaming scan of the quads matching `pat` (dictionary ids: subject,
5259    /// predicate, object) over every named graph of a quad-index file, as
5260    /// `(graph subject id, (s, p, o))`; see [`crate::quad::QuadScan`]. `None`
5261    /// for another layout, or when the section cannot be read.
5262    pub(crate) fn named_quad_scan(
5263        &self,
5264        pat: (Option<u32>, Option<u32>, Option<u32>),
5265    ) -> Option<crate::quad::QuadScan> {
5266        match &self.named_graphs {
5267            NamedGraphsSlot::Quad(q) => q.scan(pat),
5268            _ => None,
5269        }
5270    }
5271
5272    /// [`named_graph_at`](Self::named_graph_at) under a declared demand: see
5273    /// [`named_graph_name_owned_demand`](Self::named_graph_name_owned_demand).
5274    pub(crate) fn named_graph_at_demand(
5275        &self,
5276        i: usize,
5277        exhaustive: bool,
5278    ) -> Option<(&str, &GraphIndex)> {
5279        self.named_graphs.graph_at(&self.dict, i, exhaustive)
5280    }
5281
5282    /// IRIs of the named graphs in this dataset (the default graph is unnamed).
5283    /// On a lazy ranged open this walks the whole directory (headers only) —
5284    /// prefer [`named_graph_count`](Self::named_graph_count) when the number
5285    /// is all that's needed.
5286    pub fn graph_names(&self) -> Vec<&str> {
5287        (0..self.named_graphs.count())
5288            .filter_map(|i| self.named_graphs.name_at(&self.dict, i, false))
5289            .collect()
5290    }
5291
5292    /// The permutation index of a named graph, or `None` if absent.
5293    pub fn graph_index(&self, iri: &str) -> Option<&GraphIndex> {
5294        self.named_graphs.find(&self.dict, iri)
5295    }
5296
5297    /// Match a triple pattern in dictionary-ID space (subject/predicate/object
5298    /// IDs), returning integer triples — the fast path used by the BGP engine.
5299    pub fn match_ids(
5300        &self,
5301        pattern: (Option<u32>, Option<u32>, Option<u32>),
5302    ) -> Vec<(u32, u32, u32)> {
5303        self.index.match_pattern(pattern)
5304    }
5305
5306    /// All `(subject_node, object_node)` pairs for a predicate, as unified node
5307    /// IDs — no term resolution. The fast path for graph traversal.
5308    pub fn predicate_pairs(&self, predicate: &str) -> Vec<(u32, u32)> {
5309        let pid = match self.dict.predicate_id(predicate) {
5310            Some(p) => p,
5311            None => return Vec::new(),
5312        };
5313        self.index
5314            .match_pattern((None, Some(pid), None))
5315            .into_iter()
5316            .map(|(s, _p, o)| (self.dict.subject_node(s), self.dict.object_node(o)))
5317            .collect()
5318    }
5319
5320    /// Open via a [`RangeReader`], fetching only the header and the named
5321    /// section ranges — never a linear scan of the whole resource. A full query
5322    /// open touches at most 4 ranges (header, dictionary, index, pyramid-meta).
5323    pub fn open_ranged<R: RangeReader>(reader: &R) -> Result<Self, FileError> {
5324        let head = reader.read_at(0, (HEADER_LEN as u64).min(reader.len()))?;
5325        if crate::v06::open::is_v06(&head) {
5326            // The eager path reads the whole image — the rest of it, in one
5327            // more request — and a 0x06 one then opens from memory.
5328            let mut all = head;
5329            let at = all.len() as u64;
5330            if reader.len() > at {
5331                all.extend_from_slice(&reader.read_at(at, reader.len() - at)?);
5332            }
5333            return Self::open(&all);
5334        }
5335        let header = Header::from_bytes(&head)?;
5336
5337        let dict_bytes = reader.read_at(header.dictionary_offset, header.dictionary_len)?;
5338        let dict = decode_dictionary_container(&dict_bytes, header.dict_codec)?;
5339
5340        check_named_layout(&header)?;
5341        let index_bytes = reader.read_at(header.root_dir_offset, header.root_dir_len)?;
5342        let (index, _) = decode_root_index_container(&index_bytes, &header)?;
5343        let index_section_ranges = decode_index_section_ranges(
5344            &index_bytes,
5345            header.root_dir_offset,
5346            header.perms,
5347            header.index_container_sections(),
5348        )?;
5349
5350        let pyramid = PyramidSlot::Resident(if header.pyramid_meta_len > 0 {
5351            let mb = reader.read_at(header.pyramid_meta_offset, header.pyramid_meta_len)?;
5352            Some(
5353                PyramidMeta::decode(&mb)
5354                    .map_err(|_| FileError::Container("malformed pyramid meta"))?,
5355            )
5356        } else {
5357            None
5358        });
5359
5360        // Fetch the whole TEXT_INDEX section resident (this opener does one range
5361        // read per section; the lazy opener below is the one that defers it).
5362        let (ti_off, ti_len) = header.text_index_payload();
5363        let text_index_bytes = if ti_len > 0 {
5364            Some(reader.read_at(ti_off, ti_len)?)
5365        } else {
5366            None
5367        };
5368        let text_index = resident_text_index_slot(text_index_bytes.as_deref(), header.block_codec)?;
5369        // LABEL_INDEX and SCHEMA: one range read each, held resident like the
5370        // other sections of this opener.
5371        let labels = SectionSlot::Resident(header.label_index_section().and_then(|s| {
5372            let b = reader.read_at(s.offset, s.length).ok()?;
5373            crate::label_index::LabelIndex::from_section(&b, header.block_codec).ok()
5374        }));
5375        let schema = SectionSlot::Resident(read_schema_with_header(reader, &header));
5376
5377        let named_graphs = NamedGraphsSlot::resident(
5378            &header,
5379            |off, len| Ok(reader.read_at(off, len)?),
5380            || {
5381                let nb = reader.read_at(header.named_graphs_offset, header.named_graphs_len)?;
5382                decode_named_graphs(&nb, header.block_codec, header.perms)
5383            },
5384        )?;
5385
5386        // The metadata section (Dataset Card) is deliberately NOT fetched here:
5387        // a ranged query open keeps to its small range budget. Use `Rete::open`
5388        // (or a dedicated card fetch) when the card is actually needed.
5389        let tile_ranges =
5390            tile_file_ranges(&index_bytes, header.root_dir_offset, &index_section_ranges);
5391        Ok(Self {
5392            header,
5393            dict,
5394            index,
5395            index_section_ranges,
5396            tile_ranges,
5397            pyramid,
5398            text_index,
5399            labels,
5400            schema,
5401            named_graphs,
5402            metadata: Vec::new(),
5403            service_client: None,
5404            service_error: std::sync::Mutex::new(None),
5405            spill_threshold: std::sync::atomic::AtomicU64::new(u64::MAX),
5406            v06: None,
5407        })
5408    }
5409
5410    /// Open a `0x06` file whose first bytes are `first`: the header and root
5411    /// come from them; the dictionary and index are two-level directories whose
5412    /// leaves fault on demand; text, labels, schema, pyramid and named graphs
5413    /// are deferred. `budget` applies the default memory budget (the lazy
5414    /// remote path) as [`open_ranged_lazy`](Self::open_ranged_lazy) does.
5415    fn open_v06(
5416        reader: std::sync::Arc<dyn RangeReader + Send + Sync>,
5417        first: &[u8],
5418        budget: bool,
5419    ) -> Result<Self, FileError> {
5420        let f = std::sync::Arc::new(crate::v06::open::V06File::open(reader.clone(), first)?);
5421        let header = crate::v06::logical_header(&f.h);
5422        let rc = reader.concurrency();
5423        let dict = f.dictionary([true; 4])?;
5424        let index = f.index(rc)?;
5425        let (pyramid, text_index, labels, schema) = v06_slots(&f);
5426        let named_graphs = match f.quad(rc) {
5427            Some(q) => NamedGraphsSlot::Quad(q?),
5428            None => NamedGraphsSlot::Resident(Vec::new()),
5429        };
5430        let rete = Self {
5431            header,
5432            dict,
5433            index,
5434            index_section_ranges: [ByteRange::default(); NUM_PERMS],
5435            tile_ranges: Default::default(),
5436            pyramid,
5437            text_index,
5438            labels,
5439            schema,
5440            named_graphs,
5441            metadata: Vec::new(),
5442            service_client: None,
5443            service_error: std::sync::Mutex::new(None),
5444            spill_threshold: std::sync::atomic::AtomicU64::new(u64::MAX),
5445            v06: Some(f),
5446        };
5447        #[cfg(not(target_arch = "wasm32"))]
5448        if budget {
5449            rete.set_memory_budget(match crate::sysmem::default_query_budget_mb() {
5450                0 => None,
5451                mb => Some(mb << 20),
5452            });
5453        }
5454        #[cfg(target_arch = "wasm32")]
5455        let _ = budget;
5456        Ok(rete)
5457    }
5458
5459    /// The open `0x06` file (header, root, reader), `None` for a `0x05` open.
5460    pub fn v06(&self) -> Option<&crate::v06::open::V06File> {
5461        self.v06.as_deref()
5462    }
5463
5464    /// Open via an **owned** [`RangeReader`] with lazy tile faulting (tiled
5465    /// v0.2 files): fetches the header, dictionary, pyramid meta, named graphs,
5466    /// and each permutation's tile **directory** — but no default-graph tile
5467    /// payloads. Tiles fault in (one range request each) the first time a scan
5468    /// touches them, so a selective SPARQL query fetches O(touched tiles)
5469    /// bytes instead of the whole index.
5470    ///
5471    /// **Failure contract:** scans are infallible by design, so a failed tile
5472    /// fetch yields an empty tile and sets a sticky flag — after evaluating,
5473    /// callers MUST check [`index_incomplete`](Self::index_incomplete) and
5474    /// surface an error instead of the (possibly partial) results.
5475    pub fn open_ranged_lazy<R: RangeReader + Send + Sync + 'static>(
5476        reader: R,
5477    ) -> Result<Self, FileError> {
5478        Self::open_ranged_lazy_with(reader, crate::v06::FIRST_READ)
5479    }
5480
5481    /// [`open_ranged_lazy`](Self::open_ranged_lazy) with an explicit first-read
5482    /// size (at least the 1 KiB a `0x05` header needs). The default is
5483    /// [`crate::v06::FIRST_READ`]; a smaller one keeps a byte-accounting test
5484    /// of a tiny `0x05` file from reading the whole file up front.
5485    pub fn open_ranged_lazy_with<R: RangeReader + Send + Sync + 'static>(
5486        reader: R,
5487        first_read: u64,
5488    ) -> Result<Self, FileError> {
5489        // One first read — 64 KiB, or the whole file when smaller — decides the
5490        // format, and every later read inside it is served from memory.
5491        let first_len = first_read.max(HEADER_LEN as u64).min(reader.len());
5492        let first: std::sync::Arc<[u8]> = std::sync::Arc::from(reader.read_at(0, first_len)?);
5493        #[cfg(feature = "v06-parity")]
5494        if !crate::v06::open::is_v06(&first) {
5495            let all = reader.read_at(0, reader.len())?;
5496            if let Some(v6) = parity_image(&all) {
5497                let r: std::sync::Arc<dyn RangeReader + Send + Sync> =
5498                    std::sync::Arc::new(crate::reader::BytesReader(std::sync::Arc::from(v6)));
5499                let head = r.read_at(0, crate::v06::FIRST_READ.min(r.len()))?;
5500                return Self::open_v06(r, &head, true);
5501            }
5502        }
5503        let reader = crate::reader::PrefixReader::new(reader, first.clone());
5504        if crate::v06::open::is_v06(&first) {
5505            let r: std::sync::Arc<dyn RangeReader + Send + Sync> = std::sync::Arc::new(reader);
5506            return Self::open_v06(r, &first, true);
5507        }
5508        let head = reader.read_at(0, HEADER_LEN as u64)?;
5509        let header = Header::from_bytes(&head)?;
5510        check_named_layout(&header)?;
5511        let reader = std::sync::Arc::new(reader);
5512        // Captured before the loader closures take the Arc: the reader's
5513        // concurrent-range fan-out, stamped onto the index for the planner.
5514        let read_concurrency = reader.concurrency();
5515
5516        // Lazily-chunked dictionary: locate the four sections, fetch each
5517        // section's header + restart table + chunk directory (small), and
5518        // fault the chunk bodies in on first term lookup.
5519        let open_dict = || ranged_chunked_dictionary(&reader, &header, [true; 4]);
5520
5521        // Locate the six index section payloads (container framing only) and
5522        // fetch just their tile directories — shared with the per-named-graph
5523        // lazy opener, which opens a large graph's container the same way.
5524        let reader_dyn: std::sync::Arc<dyn RangeReader + Send + Sync> = reader.clone();
5525        let open_index = || {
5526            open_index_container_lazy(
5527                &reader_dyn,
5528                ByteRange {
5529                    offset: header.root_dir_offset,
5530                    len: header.root_dir_len,
5531                },
5532                header.block_codec,
5533                header.has_tile_synopsis(),
5534                read_concurrency,
5535                header.perms,
5536                header.index_container_sections(),
5537            )
5538        };
5539
5540        // Each of the two is a chain of small dependent reads (a container's
5541        // sections are found by walking their length prefixes), ~8 round trips
5542        // apiece on a real file, and the header already says where both
5543        // containers start. So a reader that can keep several reads in flight
5544        // walks them at the same time: the open costs the longer chain, not
5545        // the sum. Same reads, same result; only their overlap changes.
5546        let (dict, index) = open_side_by_side(read_concurrency, open_dict, open_index);
5547        let dict = dict?;
5548        let (index, index_section_ranges, tile_ranges) = index?;
5549
5550        // The pyramid meta is large on real graphs (tens of MB) and SPARQL never
5551        // reads it, so defer its fetch: it faults in only if `pyramid()` is
5552        // called (community / pyramid_tree / inspect queries).
5553        let pyramid = ranged_pyramid_slot(&reader, &header);
5554
5555        // The TEXT_INDEX section is also deferred: a text search faults the token
5556        // table on first call (the leading varint then its compressed bytes), then
5557        // fetches individual posting lists by `(offset, len)` — never the whole
5558        // postings blob. A SPARQL query, which never searches, pays nothing.
5559        let text_index = ranged_text_index_slot(&reader, &header);
5560
5561        // LABEL_INDEX: the first prefix search reads its head and walks down.
5562        // SCHEMA: read once, on first use — a stats question, or the planner
5563        // ordering a join that needs a measured selectivity.
5564        let labels = ranged_label_slot(&reader, &header);
5565        let schema = ranged_schema_slot(&reader, &header);
5566
5567        // Named graphs are the last eagerly-fetched section standing on this
5568        // path — and on a many-graph file they dwarf everything else (67 MB
5569        // fetched and ~32k graph indexes decoded before the first query, on a
5570        // file whose queries then touch a handful of graphs). Defer them like
5571        // the pyramid and text index: nothing is read at open; a query walks
5572        // the directory and decodes only the graphs it touches.
5573        let named_graphs = if let Some(q) = header.quad_index_section() {
5574            // The quad index: nothing read at open either; the preamble on the
5575            // first named-graph touch, then only the routing parts and tiles
5576            // a query uses.
5577            NamedGraphsSlot::Quad(crate::quad::QuadNamed::new(
5578                reader_dyn,
5579                ByteRange {
5580                    offset: q.offset,
5581                    len: q.length,
5582                },
5583                header.block_codec,
5584                header.perms,
5585                read_concurrency,
5586            ))
5587        } else if header.named_graphs_len > 0 {
5588            NamedGraphsSlot::Lazy(LazyNamedGraphs::new(
5589                reader_dyn,
5590                ByteRange {
5591                    offset: header.named_graphs_offset,
5592                    len: header.named_graphs_len,
5593                },
5594                header.block_codec,
5595                header.has_tile_synopsis(),
5596                read_concurrency,
5597                header.perms,
5598            ))
5599        } else {
5600            NamedGraphsSlot::Resident(Vec::new())
5601        };
5602
5603        let rete = Self {
5604            header,
5605            dict,
5606            index,
5607            index_section_ranges,
5608            tile_ranges,
5609            pyramid,
5610            text_index,
5611            labels,
5612            schema,
5613            named_graphs,
5614            metadata: Vec::new(),
5615            service_client: None,
5616            service_error: std::sync::Mutex::new(None),
5617            spill_threshold: std::sync::atomic::AtomicU64::new(u64::MAX),
5618            v06: None,
5619        };
5620        // A lazy open is the path for files too big to read whole, and the
5621        // one a long-lived session (relay, Python/R/Java clients, `rete serve`)
5622        // keeps resident across queries. Unbounded, its dictionary and tile
5623        // caches grow with everything any query ever touched: a GROUP BY over
5624        // a 9.5-billion-triple file ran a 16 GB container out of memory. Cap
5625        // them by default; a caller that wants another figure (or unlimited)
5626        // calls `set_memory_budget` again. The default is sized from the
5627        // memory this process can still use (`RETE_QUERY_BUDGET_MB` overrides
5628        // it; see `crate::sysmem`): a fixed 4096 MiB did not fit a 4 GiB
5629        // container. Not on wasm32, whose 4 GiB address space makes the default
5630        // meaningless and whose callers bound their own range budgets.
5631        #[cfg(not(target_arch = "wasm32"))]
5632        rete.set_memory_budget(match crate::sysmem::default_query_budget_mb() {
5633            0 => None,
5634            mb => Some(mb << 20),
5635        });
5636        Ok(rete)
5637    }
5638
5639    /// Did any lazy fetch (index tile or dictionary chunk) fail since this
5640    /// `Rete` was opened? When true, query results may be silently incomplete —
5641    /// callers using [`Rete::open_ranged_lazy`] must check this after
5642    /// evaluating and turn it into an error.
5643    pub fn index_incomplete(&self) -> bool {
5644        self.index.load_incomplete()
5645            || self.dict.load_incomplete()
5646            || self.named_graphs.load_incomplete()
5647    }
5648
5649    /// Forget recorded lazy-fetch failures — the start-of-evaluation reset for
5650    /// a RESIDENT session (a browser worker holding one `Rete` across many
5651    /// queries): it makes [`index_incomplete`](Self::index_incomplete) a
5652    /// per-query verdict instead of a per-open one, so a single transient
5653    /// network failure no longer fails every subsequent query on the session.
5654    /// Sound because failed tiles/chunks are never cached — the next
5655    /// evaluation simply retries the fetch.
5656    pub fn reset_load_failures(&self) {
5657        self.index.reset_load_failure();
5658        self.dict.reset_load_failure();
5659        self.named_graphs.reset_load_failures();
5660    }
5661
5662    /// Bound peak resident memory to roughly `budget` bytes by capping the
5663    /// evictable dictionary chunk cache and the default-graph index tile cache
5664    /// (bounded-export phase 2). `None` or `u64::MAX` means unlimited — nothing
5665    /// is evicted and residency is identical to a plain open, which is what
5666    /// `--in-memory` and every query/wasm path use.
5667    ///
5668    /// Sound at any cap: the dictionary and the default-graph index both fault
5669    /// their chunks/tiles through a loader, so an evicted body is simply
5670    /// re-faulted on its next touch; the export dump reads each needed body
5671    /// while holding it (the cache `Arc` handout invariant), so eviction never
5672    /// drops bytes in use. Named-graph indexes are **not** capped here — a small
5673    /// resident graph has no loader and could never reload an evicted tile — so
5674    /// they stay unlimited and are released whole per slot instead (see
5675    /// [`release_named_graph`](Self::release_named_graph)).
5676    ///
5677    /// The open's **fixed** heap — chunk and tile directories, the named-graph
5678    /// directory walked so far ([`fixed_overhead`](Self::fixed_overhead)) — is
5679    /// paid whatever the caches hold, so it comes out of the dictionary share:
5680    /// the caps plus the fixed heap stay within `budget`. It grows with the
5681    /// file's chunk/tile/graph count (~18 MB of directories on the 14.75 GiB
5682    /// openaire-relation, 88 MB with its tile directories; 224 MB on a pre-#198
5683    /// epfl-infoscience whose chunk keys are whole boundary terms). Calling
5684    /// this again after a walk of the named-graph directory (as `rete export`
5685    /// does once it has listed the graph slots) re-subtracts the larger figure.
5686    /// The dictionary share never drops below its 64 MiB floor, which is the
5687    /// only way the total can be exceeded: a file whose directories alone
5688    /// approach the budget.
5689    ///
5690    /// Returns the split actually applied (dictionary share net of the fixed
5691    /// heap), for `RETE_OPEN_DEBUG` logging.
5692    pub fn set_memory_budget(&self, budget: Option<u64>) -> MemoryBudget {
5693        let total = budget.unwrap_or(u64::MAX);
5694        self.set_spill_threshold(budget.map(spill_threshold_for));
5695        let mut split = split_memory_budget(total);
5696        let fixed = if total == u64::MAX {
5697            0
5698        } else {
5699            self.fixed_overhead().total()
5700        };
5701        if split.dict_cache != u64::MAX {
5702            split.dict_cache = split
5703                .dict_cache
5704                .saturating_sub(fixed)
5705                .max(MIN_DICT_CACHE_CAP);
5706        }
5707        self.dict.set_cache_cap(split.dict_cache);
5708        self.index.set_cache_cap(split.tile_cache);
5709        // Large named graphs (opened lazily, loader-bearing) get the same tile
5710        // cap; small resident ones stay unlimited (bounded by size + per-slot
5711        // release). Combined with `release_named_graph`, a many-graph dump holds
5712        // one graph's tiles at a time, each capped.
5713        match &self.named_graphs {
5714            NamedGraphsSlot::Lazy(lazy) => lazy.set_tile_cap(split.tile_cache),
5715            // Every quad-index view, and the tiles they share, re-fault
5716            // through loaders — on every open path — so they take the cap too.
5717            NamedGraphsSlot::Quad(q) => q.set_tile_cap(split.tile_cache),
5718            NamedGraphsSlot::Resident(_) => {}
5719        }
5720        if std::env::var("RETE_OPEN_DEBUG").is_ok() {
5721            eprintln!(
5722                "[memory-budget] total={} dict_cache={} tile_cache={} block_cache={} fixed={:.1}MiB",
5723                if total == u64::MAX {
5724                    "unlimited".to_string()
5725                } else {
5726                    format!("{}MiB", total >> 20)
5727                },
5728                fmt_cap(split.dict_cache),
5729                fmt_cap(split.tile_cache),
5730                fmt_cap(split.block_cache),
5731                fixed as f64 / (1u64 << 20) as f64,
5732            );
5733        }
5734        split
5735    }
5736
5737    /// How many bytes a query's `COUNT(DISTINCT …)` keeps in memory before it
5738    /// writes sorted runs to the temporary directory (`TMPDIR`) and counts by
5739    /// merging them; the count is exact either way. `u64::MAX` (the default of
5740    /// an eager open) never spills. [`set_memory_budget`](Self::set_memory_budget)
5741    /// sets it to an eighth of the budget, so a lazy open spills past 512 MiB
5742    /// by default.
5743    pub fn spill_threshold(&self) -> u64 {
5744        self.spill_threshold
5745            .load(std::sync::atomic::Ordering::Relaxed)
5746    }
5747
5748    /// Set [`spill_threshold`](Self::spill_threshold) without touching the
5749    /// caches — for an eager (resident) open, whose caches cannot be capped.
5750    /// `None` = never spill.
5751    pub fn set_spill_threshold(&self, bytes: Option<u64>) {
5752        self.spill_threshold.store(
5753            bytes.unwrap_or(u64::MAX),
5754            std::sync::atomic::Ordering::Relaxed,
5755        );
5756    }
5757
5758    /// The dictionary chunk cache's observability counters (decoded chunks and
5759    /// bytes, hits/misses, evictions, resident bytes) — for the bounded-export
5760    /// work-accounting tests and `RETE_OPEN_DEBUG`.
5761    pub fn dict_cache_stats(&self) -> crate::chunk_cache::CacheStats {
5762        self.dict.cache_stats()
5763    }
5764
5765    /// Heap bytes this open holds outside the bounded caches — the dictionary
5766    /// chunk directories, the index tile directories and the named-graph
5767    /// directory — for `RETE_OPEN_DEBUG` / `RETE_MEM_TRACE` accounting.
5768    pub fn fixed_overhead(&self) -> FixedOverhead {
5769        FixedOverhead {
5770            dict_directory: self.dict.directory_bytes(),
5771            index_directory: self.index.directory_bytes(),
5772            named_graphs: self.named_graphs.directory_bytes(),
5773        }
5774    }
5775
5776    /// The default-graph index tile cache's observability counters.
5777    pub fn index_cache_stats(&self) -> crate::chunk_cache::CacheStats {
5778        self.index.cache_stats()
5779    }
5780
5781    /// The current dictionary chunk-cache byte cap (`u64::MAX` = unlimited).
5782    pub fn dict_cache_cap(&self) -> u64 {
5783        self.dict.cache_cap()
5784    }
5785
5786    /// Release a named graph's decoded index after its slot has been exported,
5787    /// dropping its `GraphIndex` (and, with it, that graph's decoded tiles and
5788    /// its own tile cache) so a many-graph dump does not accumulate one resident
5789    /// index per graph — the next linear floor on many-graph files once the
5790    /// dictionary is bounded (plan §6). A no-op for the default graph, an
5791    /// unknown IRI, a resident (in-memory/eager) open, or a graph never opened.
5792    ///
5793    /// Takes `&mut self`: the export loop finishes one slot's dump (which only
5794    /// borrows `&self`) before releasing, so no outstanding borrow of the index
5795    /// exists. Correct because a released index is simply re-opened (headers +
5796    /// tile directory) on a later access — the same lazy contract as a chunk
5797    /// that was never faulted.
5798    pub fn release_named_graph(&mut self, iri: &str) {
5799        match &mut self.named_graphs {
5800            NamedGraphsSlot::Lazy(lazy) => lazy.release(iri),
5801            NamedGraphsSlot::Quad(q) => q.release(&self.dict, iri),
5802            NamedGraphsSlot::Resident(_) => {}
5803        }
5804    }
5805
5806    fn resolve_query_pattern(
5807        &self,
5808        s: Option<&str>,
5809        p: Option<&str>,
5810        o: Option<&str>,
5811    ) -> Option<Pattern> {
5812        let sid = match s {
5813            Some(t) => match self.dict.subject_id(t) {
5814                Some(id) => Some(id),
5815                None => return None,
5816            },
5817            None => None,
5818        };
5819        let pid = match p {
5820            Some(t) => match self.dict.predicate_id(t) {
5821                Some(id) => Some(id),
5822                None => return None,
5823            },
5824            None => None,
5825        };
5826        let oid = match o {
5827            Some(t) => match self.dict.object_id(t) {
5828                Some(id) => Some(id),
5829                None => return None,
5830            },
5831            None => None,
5832        };
5833        Some((sid, pid, oid))
5834    }
5835
5836    /// Evaluate a triple pattern and include the file/index provenance for every
5837    /// matched result. A bound term that is unknown to the dictionary yields no
5838    /// matches.
5839    pub fn query_with_provenance(
5840        &self,
5841        s: Option<&str>,
5842        p: Option<&str>,
5843        o: Option<&str>,
5844    ) -> Vec<TripleProvenance> {
5845        let pattern = match self.resolve_query_pattern(s, p, o) {
5846            Some(pattern) => pattern,
5847            None => return Vec::new(),
5848        };
5849
5850        let index_permutation = GraphIndex::best_permutation_in(self.header.perms, pattern);
5851        let si = index_permutation.section_index();
5852        let (dictionary_range, index_range, index_section_range, pyramid_range) = match &self.v06 {
5853            Some(f) => {
5854                use crate::v06::kind;
5855                let sec = |k| f.section(k).unwrap_or_default();
5856                let perm = f
5857                    .root
5858                    .unit(crate::v06::tag::INDEX)
5859                    .and_then(|u| crate::v06::decode_index_root(u).ok())
5860                    .and_then(|v| v.into_iter().find(|p| p.si == si))
5861                    .map(|p| ByteRange {
5862                        offset: sec(kind::INDEX_DATA).offset + p.data_off,
5863                        len: p.data_len,
5864                    })
5865                    .unwrap_or_default();
5866                (
5867                    sec(kind::DICT_DATA),
5868                    sec(kind::INDEX_DATA),
5869                    perm,
5870                    f.section(kind::PYRAMID),
5871                )
5872            }
5873            None => (
5874                ByteRange {
5875                    offset: self.header.dictionary_offset,
5876                    len: self.header.dictionary_len,
5877                },
5878                ByteRange {
5879                    offset: self.header.root_dir_offset,
5880                    len: self.header.root_dir_len,
5881                },
5882                self.index_section_ranges[si],
5883                (self.header.pyramid_meta_len > 0).then_some(ByteRange {
5884                    offset: self.header.pyramid_meta_offset,
5885                    len: self.header.pyramid_meta_len,
5886                }),
5887            ),
5888        };
5889
5890        let tiles = &self.tile_ranges[si];
5891        let v06_tiles = self.v06.is_some().then(|| self.index.tile_sections()[si]);
5892        self.index
5893            .match_pattern(pattern)
5894            .into_iter()
5895            .filter_map(|(s, p, o)| {
5896                let terms = (
5897                    self.dict.subject_term(s)?,
5898                    self.dict.predicate_term(p)?,
5899                    self.dict.object_term(o)?,
5900                );
5901                // The physical tile holding this match: the one whose
5902                // leading-id range covers the match's permuted leading id.
5903                let a = index_permutation.forward((s, p, o)).0;
5904                let (tile, tile_range) = if let Some(tl) = v06_tiles {
5905                    let ti = tl.first_max_ge(a);
5906                    match (ti < tl.len() && tl[ti].leading_range().0 <= a)
5907                        .then(|| tl.range(ti))
5908                        .flatten()
5909                    {
5910                        Some((offset, len)) => (
5911                            Some(format!("{}/{ti}", index_permutation.name())),
5912                            Some(ByteRange { offset, len }),
5913                        ),
5914                        None => (None, None),
5915                    }
5916                } else {
5917                    let ti = tiles.partition_point(|&(_, max_a, _)| max_a < a);
5918                    match tiles.get(ti) {
5919                        Some(&(min_a, _, range)) if min_a <= a => (
5920                            Some(format!("{}/{ti}", index_permutation.name())),
5921                            Some(range),
5922                        ),
5923                        _ => (None, None),
5924                    }
5925                };
5926                Some(TripleProvenance {
5927                    terms,
5928                    ids: (s, p, o),
5929                    graph: None,
5930                    matched_pattern: pattern,
5931                    index_permutation,
5932                    dictionary_range,
5933                    index_range,
5934                    index_section_range,
5935                    pyramid_range,
5936                    tile,
5937                    tile_range,
5938                })
5939            })
5940            .collect()
5941    }
5942
5943    /// Evaluate a triple pattern given as optional term strings, returning
5944    /// matching triples resolved back to terms. A bound term that is unknown to
5945    /// the dictionary yields no matches.
5946    pub fn query(&self, s: Option<&str>, p: Option<&str>, o: Option<&str>) -> Vec<TermTriple> {
5947        self.query_with_provenance(s, p, o)
5948            .into_iter()
5949            .map(|m| m.terms)
5950            .collect()
5951    }
5952
5953    /// Match a triple pattern **within a single graph** — `None` is the default
5954    /// graph, `Some(iri)` a named graph — resolving matches to canonical terms.
5955    /// This is [`Rete::query`] (default-graph only) generalized to any graph: the
5956    /// graph-scoped primitive a quad-aware consumer (e.g. an RDF4J `Sail`'s
5957    /// `getStatements`) needs. An unknown graph IRI, or a bound term absent from
5958    /// the shared dictionary, yields an empty result. All graphs share one
5959    /// dictionary, so the pattern resolves once against that ID space.
5960    pub fn query_in_graph(
5961        &self,
5962        graph: Option<&str>,
5963        s: Option<&str>,
5964        p: Option<&str>,
5965        o: Option<&str>,
5966    ) -> Vec<TermTriple> {
5967        let pattern = match self.resolve_query_pattern(s, p, o) {
5968            Some(pattern) => pattern,
5969            None => return Vec::new(),
5970        };
5971        let index = match graph {
5972            None => &self.index,
5973            Some(g) => match self.graph_index(g) {
5974                Some(i) => i,
5975                None => return Vec::new(),
5976            },
5977        };
5978        self.resolve_matches(index.match_pattern(pattern))
5979    }
5980
5981    /// Resolve matched id-triples to terms, with one coalesced dictionary
5982    /// fault for exactly their terms.
5983    fn resolve_matches(&self, ids: Vec<Triple>) -> Vec<TermTriple> {
5984        // One coalesced dictionary fault for THIS pattern's output, not the
5985        // whole dictionary. `prefetch_all` used to sit here, which made a
5986        // bounded `(s, p, o, g)` lookup — the primitive an RDF4J `Sail`'s
5987        // `getStatements` calls, once per named graph via `query_quads` — pay
5988        // for every term in the file: on `cordis.rete` (801 MB, a 417 MB
5989        // dictionary) a 15-row answer read 416 MB and peaked at 2.5 GB RSS.
5990        // Same call `dump_batch` makes, for the same reason.
5991        if !ids.is_empty() {
5992            let mut nodes = Vec::with_capacity(ids.len() * 2);
5993            let mut preds = Vec::with_capacity(ids.len());
5994            for &(s, p, o) in &ids {
5995                nodes.push(self.dict.subject_node(s));
5996                nodes.push(self.dict.object_node(o));
5997                preds.push(p);
5998            }
5999            self.dict.prefetch_terms(&nodes, &preds);
6000        }
6001        ids.into_iter()
6002            .filter_map(|(s, p, o)| {
6003                Some((
6004                    self.dict.subject_term(s)?,
6005                    self.dict.predicate_term(p)?,
6006                    self.dict.object_term(o)?,
6007                ))
6008            })
6009            .collect()
6010    }
6011
6012    /// Match a triple pattern across the default graph **and every named graph**,
6013    /// tagging each match with its graph (`None` = default). The quad-level
6014    /// companion to [`Rete::query`]; default-graph matches come first, then each
6015    /// named graph in stored order.
6016    pub fn query_quads(
6017        &self,
6018        s: Option<&str>,
6019        p: Option<&str>,
6020        o: Option<&str>,
6021    ) -> Vec<(TermTriple, Option<String>)> {
6022        let mut out: Vec<(TermTriple, Option<String>)> = self
6023            .query_in_graph(None, s, p, o)
6024            .into_iter()
6025            .map(|t| (t, None))
6026            .collect();
6027        let Some(pattern) = self.resolve_query_pattern(s, p, o) else {
6028            return out;
6029        };
6030        for i in 0..self.named_graphs.count() {
6031            // A quad dump visits every graph by construction — exhaustive
6032            // demand, so a lazy walk may read the section in bulk chunks.
6033            let Some(iri) = self.named_graphs.name_at(&self.dict, i, true) else {
6034                continue;
6035            };
6036            // A small quad-index graph through an owned view, dropped here:
6037            // the walk holds one view at a time, not one per graph.
6038            let owned = self.named_graph_transient(i);
6039            let index = match &owned {
6040                Some(view) => view,
6041                None => match self.named_graphs.graph_at(&self.dict, i, true) {
6042                    Some((_, g)) => g,
6043                    None => continue,
6044                },
6045            };
6046            for triple in self.resolve_matches(index.match_pattern(pattern)) {
6047                out.push((triple, Some(iri.to_string())));
6048            }
6049        }
6050        out
6051    }
6052
6053    /// Evaluate one triple pattern through a [`RangeReader`] by fetching only
6054    /// the header, the dictionary, and — for a tiled (v0.2) file — the
6055    /// selected permutation section's tile **directory** plus the tile(s) the
6056    /// bound leading id routes to; an unbound leading id fetches the section's
6057    /// tile body in one request. v0.1 files fetch the whole selected section.
6058    /// Unknown bound terms return an empty result before touching the index.
6059    pub fn query_ranged<R: RangeReader>(
6060        reader: &R,
6061        s: Option<&str>,
6062        p: Option<&str>,
6063        o: Option<&str>,
6064    ) -> Result<Vec<TermTriple>, FileError> {
6065        let head = reader.read_at(0, (HEADER_LEN as u64).min(reader.len()))?;
6066        if crate::v06::open::is_v06(&head) {
6067            return Ok(v06_query_ranged(reader, head, s, p, o, true)?.unwrap_or_default());
6068        }
6069        let routed = match route_pattern(reader, head, s, p, o)? {
6070            Some(routed) => routed,
6071            None => return Ok(Vec::new()),
6072        };
6073        let matches = fetch_routed_matches(reader, &routed)?;
6074        Ok(matches
6075            .into_iter()
6076            .filter_map(|(s, p, o)| {
6077                Some((
6078                    routed.dict.subject_term(s)?,
6079                    routed.dict.predicate_term(p)?,
6080                    routed.dict.object_term(o)?,
6081                ))
6082            })
6083            .collect())
6084    }
6085
6086    /// Route one triple pattern to its permutation section without fetching
6087    /// any payload bytes. Returns `false` when a bound term is unknown and the
6088    /// index was skipped.
6089    pub fn route_pattern_ranged<R: RangeReader>(
6090        reader: &R,
6091        s: Option<&str>,
6092        p: Option<&str>,
6093        o: Option<&str>,
6094    ) -> Result<bool, FileError> {
6095        let head = reader.read_at(0, (HEADER_LEN as u64).min(reader.len()))?;
6096        if crate::v06::open::is_v06(&head) {
6097            return Ok(v06_query_ranged(reader, head, s, p, o, false)?.is_some());
6098        }
6099        Ok(route_pattern(reader, head, s, p, o)?.is_some())
6100    }
6101}
6102
6103/// [`Rete::query_ranged`] / [`Rete::route_pattern_ranged`] on a `0x06` file
6104/// through a borrowed reader: the front and the dictionary data — contiguous,
6105/// what the `0x05` path reads as header + dictionary — in one more request
6106/// after `head`, the pattern resolved there, then (with `fetch`) the routed
6107/// tiles in one span. `None` when a bound term is not in the dictionary.
6108fn v06_query_ranged<R: RangeReader>(
6109    reader: &R,
6110    head: Vec<u8>,
6111    s: Option<&str>,
6112    p: Option<&str>,
6113    o: Option<&str>,
6114    fetch: bool,
6115) -> Result<Option<Vec<TermTriple>>, FileError> {
6116    use crate::v06::{kind, open::SparseReader, Header06, Root06, HEADER_LEN};
6117    let len = reader.len();
6118    let h = Header06::from_bytes(&head)?;
6119    let root_end = (HEADER_LEN as u64 + h.root_len as u64).min(len);
6120    let root_bytes = if root_end <= head.len() as u64 {
6121        head[HEADER_LEN..root_end as usize].to_vec()
6122    } else {
6123        reader.read_at(HEADER_LEN as u64, root_end - HEADER_LEN as u64)?
6124    };
6125    let root = Root06::decode(&root_bytes, h.internal_codec)?;
6126    let dict = root.section(kind::DICT_DATA).and_then(|d| d.abs(&h).ok());
6127    // The open below reads only the 1 KiB it must (not the 64 KiB first
6128    // read): what it needs past that is the front and the dictionary, here.
6129    let end = dict
6130        .map_or(h.data_offset, |d| d.offset + d.len)
6131        .max(crate::header::HEADER_LEN as u64)
6132        .min(len);
6133    let mut held = head;
6134    if end > held.len() as u64 {
6135        let at = held.len() as u64;
6136        held.extend_from_slice(&reader.read_at(at, end - at)?);
6137    }
6138    let sparse = SparseReader::fetch(reader, held, &[])?;
6139    let rete = Rete::open_ranged_lazy_with(sparse, 0)?;
6140    let Some(pattern) = resolve_query_pattern(rete.dictionary(), s, p, o) else {
6141        return Ok(None);
6142    };
6143    if !fetch {
6144        return Ok(Some(Vec::new()));
6145    }
6146    let perm = GraphIndex::best_permutation_in(h.perms(), pattern);
6147    let si = perm.section_index();
6148    let index = rete.default_index();
6149    let list = index.tile_sections()[si];
6150    let n = list.len();
6151    let [pa, _, _] = perm.order_pattern(pattern);
6152    let (start, stop) = match pa {
6153        Some(a) => {
6154            let start = list.first_max_ge(a);
6155            let mut stop = start;
6156            while stop < n {
6157                list.ensure_span(stop, stop + 1);
6158                if list[stop].leading_range().0 > a {
6159                    break;
6160                }
6161                stop += 1;
6162            }
6163            (start, stop)
6164        }
6165        None => {
6166            list.ensure_span(0, n);
6167            (0, n)
6168        }
6169    };
6170    let ranges: Vec<(u64, u64)> = (start..stop)
6171        .filter_map(|i| index.tile_range(si, i))
6172        .collect();
6173    let mut out = Vec::new();
6174    if let (Some(&(lo, _)), Some(&(last, last_len))) = (ranges.first(), ranges.last()) {
6175        // The routed tiles are one contiguous run of the section: one request.
6176        let body = reader.read_at(lo, last + last_len - lo)?;
6177        for &(off, l) in &ranges {
6178            let at = (off - lo) as usize;
6179            let tile = decompress(h.block_codec, &body[at..at + l as usize])?;
6180            out.extend(GraphIndex::match_serialized_block(&tile, perm, pattern));
6181        }
6182    }
6183    out.sort_unstable();
6184    let d = rete.dictionary();
6185    Ok(Some(
6186        out.into_iter()
6187            .filter_map(|(s, p, o)| {
6188                Some((d.subject_term(s)?, d.predicate_term(p)?, d.object_term(o)?))
6189            })
6190            .collect(),
6191    ))
6192}
6193
6194/// A pattern routed to its permutation section: everything needed to fetch
6195/// matches, with no payload bytes read yet.
6196struct RoutedPattern {
6197    dict: Dictionary,
6198    pattern: Pattern,
6199    permutation: IndexPermutation,
6200    header: Header,
6201    /// Absolute byte range of the selected section's payload.
6202    section: ByteRange,
6203}
6204
6205/// Resolve a pattern against the remote dictionary and locate its permutation
6206/// section (header + dictionary + container framing only).
6207fn route_pattern<R: RangeReader>(
6208    reader: &R,
6209    head: Vec<u8>,
6210    s: Option<&str>,
6211    p: Option<&str>,
6212    o: Option<&str>,
6213) -> Result<Option<RoutedPattern>, FileError> {
6214    let header = Header::from_bytes(&head)?;
6215
6216    let dict_bytes = reader.read_at(header.dictionary_offset, header.dictionary_len)?;
6217    let dict = decode_dictionary_container(&dict_bytes, header.dict_codec)?;
6218
6219    let Some(pattern) = resolve_query_pattern(&dict, s, p, o) else {
6220        return Ok(None);
6221    };
6222    let permutation = GraphIndex::best_permutation_in(header.perms, pattern);
6223    let section = locate_container_section_ranged(
6224        reader,
6225        header.root_dir_offset,
6226        header.root_dir_len,
6227        header
6228            .perms
6229            .position(permutation)
6230            .ok_or(FileError::Container("routed to an absent permutation"))?,
6231        header.index_container_sections() as u64,
6232    )?;
6233    Ok(Some(RoutedPattern {
6234        dict,
6235        pattern,
6236        permutation,
6237        header,
6238        section,
6239    }))
6240}
6241
6242/// Fetch and scan a routed pattern's matches: read the tile directory, then only
6243/// the matching tile byte ranges (the run of covering tiles for a bound leading
6244/// id — one for ordinary groups, several for a split mega-group — the
6245/// O(matching bytes) promise).
6246fn fetch_routed_matches<R: RangeReader>(
6247    reader: &R,
6248    routed: &RoutedPattern,
6249) -> Result<Vec<Triple>, FileError> {
6250    let dir = read_tile_directory_ranged(reader, routed.section)?;
6251    let [pa, _, _] = routed.permutation.order_pattern(routed.pattern);
6252    let codec = routed.header.block_codec;
6253    let mut out = Vec::new();
6254    match pa {
6255        // Bound leading id: the run of covering tiles (several for a split
6256        // mega-group; one otherwise).
6257        Some(a) => {
6258            for e in dir.iter().filter(|e| e.min_a <= a && a <= e.max_a) {
6259                let len = e.end - e.start;
6260                let bytes = reader.read_at(checked_end(routed.section.offset, e.start)?, len)?;
6261                if bytes.len() as u64 != len {
6262                    return Err(FileError::Container("short tile read"));
6263                }
6264                let tile = decompress(codec, &bytes)?;
6265                out.extend(GraphIndex::match_serialized_block(
6266                    &tile,
6267                    routed.permutation,
6268                    routed.pattern,
6269                ));
6270            }
6271        }
6272        // Unbound leading id: every tile matters — fetch the contiguous tile
6273        // body in one request and slice it.
6274        None => {
6275            if let (Some(first), Some(last)) = (dir.first(), dir.last()) {
6276                let base = first.start;
6277                let body =
6278                    reader.read_at(checked_end(routed.section.offset, base)?, last.end - base)?;
6279                for e in &dir {
6280                    // Checked: a reader may return fewer bytes than asked.
6281                    let stored = checked_slice(&body, e.start - base, e.end - e.start)
6282                        .ok_or(FileError::Container("short tile read"))?;
6283                    let tile = decompress(codec, stored)?;
6284                    out.extend(GraphIndex::match_serialized_block(
6285                        &tile,
6286                        routed.permutation,
6287                        routed.pattern,
6288                    ));
6289                }
6290            }
6291        }
6292    }
6293    out.sort_unstable();
6294    Ok(out)
6295}
6296
6297fn resolve_query_pattern(
6298    dict: &Dictionary,
6299    s: Option<&str>,
6300    p: Option<&str>,
6301    o: Option<&str>,
6302) -> Option<Pattern> {
6303    let sid = match s {
6304        Some(t) => Some(dict.subject_id(t)?),
6305        None => None,
6306    };
6307    let pid = match p {
6308        Some(t) => Some(dict.predicate_id(t)?),
6309        None => None,
6310    };
6311    let oid = match o {
6312        Some(t) => Some(dict.object_id(t)?),
6313        None => None,
6314    };
6315    Some((sid, pid, oid))
6316}
6317
6318/// The lazily-chunked dictionary of a ranged open: per section, fetch the
6319/// header + chunk directory and fault the chunk bodies in on first term lookup.
6320///
6321/// `want` selects which of the four sections (0 shared, 1 subject-only,
6322/// 2 object-only, 3 predicates) are really read; a skipped one becomes an empty
6323/// section whose terms resolve to `None`. On a file written before #198 a
6324/// **directory is not small**: it carried each chunk's first term verbatim, so
6325/// the object-only section of a dataset that stores abstracts can run to
6326/// hundreds of megabytes, and fetching it is most of what a remote open costs.
6327/// Current builds key it by the shortest separator instead (a few bytes per
6328/// chunk), but that is write-side, so every already-published file still pays.
6329/// A reader that only ever resolves subjects (see [`SearchView`]) skips
6330/// sections 2 and 3 and pays none of it. [`Rete::open_ranged_lazy`] wants all
6331/// four.
6332fn ranged_chunked_dictionary<R: RangeReader + Send + Sync + 'static>(
6333    reader: &std::sync::Arc<R>,
6334    header: &Header,
6335    want: [bool; 4],
6336) -> Result<Dictionary, FileError> {
6337    // One (unlimited) cache shared across the four sections, keyed by
6338    // (section_index, chunk_index).
6339    let cache = crate::chunk_cache::ChunkCache::unlimited_arc();
6340    let mut dict_sections: Vec<crate::dict::ChunkedSection> = Vec::with_capacity(4);
6341    let mut section_ranges: [std::sync::Arc<Vec<ByteRange>>; 4] = Default::default();
6342    for si in 0..4 {
6343        if !want[si] {
6344            dict_sections.push(crate::dict::ChunkedSection::from_parts(
6345                crate::dict::SectionMeta {
6346                    term_count: 0,
6347                    restart_interval: 1,
6348                    restart_offsets: Vec::new(),
6349                },
6350                Vec::new(),
6351                None,
6352                cache.clone(),
6353                si as u8,
6354            ));
6355            continue;
6356        }
6357        let section = locate_container_section_ranged(
6358            reader.as_ref(),
6359            header.dictionary_offset,
6360            header.dictionary_len,
6361            si,
6362            4,
6363        )?;
6364        let (meta, entries) = read_dict_dir_ranged(reader.as_ref(), section)?;
6365        // One copy of the chunk ranges, shared by this section's loaders and
6366        // the dictionary's cross-section batch loader.
6367        let ranges: std::sync::Arc<Vec<ByteRange>> = std::sync::Arc::new(
6368            entries
6369                .iter()
6370                .map(|e| ByteRange {
6371                    offset: section.offset + e.start,
6372                    len: (e.end - e.start),
6373                })
6374                .collect(),
6375        );
6376        section_ranges[si] = ranges.clone();
6377        let chunks: Vec<crate::dict::SectionChunk> = entries
6378            .into_iter()
6379            .map(|e| crate::dict::SectionChunk::new(e.first_run, e.key, e.body_start))
6380            .collect();
6381        let chunk_reader = reader.clone();
6382        let codec = header.dict_codec;
6383        let loader_ranges = ranges.clone();
6384        let loader: crate::dict::ChunkLoader = Box::new(move |ci| {
6385            let range = loader_ranges.get(ci)?;
6386            let bytes = chunk_reader.read_at(range.offset, range.len).ok()?;
6387            decompress(codec, &bytes).ok()
6388        });
6389        // Full-section sweeps (export/dump) batch their chunk fetches:
6390        // adjacent chunk ranges coalesce into a handful of range reads.
6391        let bulk_reader = reader.clone();
6392        let bulk: crate::dict::ChunkBulkLoader = Box::new(move |cis| {
6393            let want: Option<Vec<ByteRange>> =
6394                cis.iter().map(|&ci| ranges.get(ci).copied()).collect();
6395            let blobs = read_coalesced(bulk_reader.as_ref(), &want?, DICT_COALESCE_GAP)?;
6396            blobs.iter().map(|b| decompress(codec, b).ok()).collect()
6397        });
6398        dict_sections.push(
6399            crate::dict::ChunkedSection::from_parts(
6400                meta,
6401                chunks,
6402                Some(loader),
6403                cache.clone(),
6404                si as u8,
6405            )
6406            .with_bulk_loader(bulk),
6407        );
6408    }
6409    let dict_arr: [crate::dict::ChunkedSection; 4] = dict_sections
6410        .try_into()
6411        .map_err(|_| FileError::Container("expected 4 dictionary sections"))?;
6412    // Decoding a result touches several sections at once; this fetches all
6413    // their missing chunks in one batch (pairs arrive section by section in
6414    // ascending chunk order, so ascending in the file, as coalescing wants).
6415    let multi_reader = reader.clone();
6416    let codec = header.dict_codec;
6417    let multi: crate::dict::MultiChunkBulkLoader = Box::new(move |pairs| {
6418        let want: Option<Vec<ByteRange>> = pairs
6419            .iter()
6420            .map(|&(si, ci)| section_ranges.get(si as usize)?.get(ci).copied())
6421            .collect();
6422        let blobs = read_coalesced(multi_reader.as_ref(), &want?, DICT_COALESCE_GAP)?;
6423        blobs.iter().map(|b| decompress(codec, b).ok()).collect()
6424    });
6425    Ok(Dictionary::from_chunked_sections(dict_arr).with_multi_bulk_loader(multi))
6426}
6427
6428/// Run two independent steps of an open: at the same time when the reader can
6429/// keep several reads in flight (`concurrency > 1`) on a target with threads,
6430/// else one after the other. Either way both run to completion and their
6431/// results come back as they would inline; a panic in either propagates.
6432#[cfg(not(target_arch = "wasm32"))]
6433fn open_side_by_side<A, B: Send>(
6434    concurrency: usize,
6435    a: impl FnOnce() -> A,
6436    b: impl FnOnce() -> B + Send,
6437) -> (A, B) {
6438    if concurrency <= 1 {
6439        return (a(), b());
6440    }
6441    std::thread::scope(|scope| {
6442        let hb = scope.spawn(b);
6443        let ra = a();
6444        let rb = hb
6445            .join()
6446            .unwrap_or_else(|panic| std::panic::resume_unwind(panic));
6447        (ra, rb)
6448    })
6449}
6450
6451/// wasm32 has no threads here: always one after the other.
6452#[cfg(target_arch = "wasm32")]
6453fn open_side_by_side<A, B>(
6454    _concurrency: usize,
6455    a: impl FnOnce() -> A,
6456    b: impl FnOnce() -> B,
6457) -> (A, B) {
6458    (a(), b())
6459}
6460
6461/// The `v06-parity` test hook: a 0x05 image converted to 0x06 (`None` for a
6462/// 0x06 image, or one the converter refuses, such as per-graph named graphs).
6463#[cfg(feature = "v06-parity")]
6464fn parity_image(bytes: &[u8]) -> Option<Vec<u8>> {
6465    if crate::v06::open::is_v06(bytes) || bytes.len() < HEADER_LEN {
6466        return None;
6467    }
6468    crate::v06::write::convert_image(bytes, crate::v06::write::WriteOpts::default()).ok()
6469}
6470
6471/// [`Rete::file_layout`] for a `0x06` file: the header, the root, the card
6472/// region, then every front and data section, in file order.
6473fn v06_layout(f: &crate::v06::open::V06File) -> Vec<LayoutSegment> {
6474    let h = &f.h;
6475    let seg = |kind: &'static str, label: String, offset: u64, len: u64| LayoutSegment {
6476        kind,
6477        label,
6478        offset,
6479        len,
6480    };
6481    let mut out = vec![
6482        seg(
6483            "header",
6484            format!("header (fixed {} bytes)", crate::v06::HEADER_LEN),
6485            0,
6486            crate::v06::HEADER_LEN as u64,
6487        ),
6488        seg(
6489            "directory",
6490            "root block (section table, directory roots)".into(),
6491            h.root_offset(),
6492            h.root_len as u64,
6493        ),
6494    ];
6495    let c = h.card_range();
6496    if c.len > 0 {
6497        out.push(seg(
6498            "metadata",
6499            "metadata (dataset card, zstd)".into(),
6500            c.offset,
6501            c.len,
6502        ));
6503    }
6504    let b = h.build_range();
6505    if b.len > 0 {
6506        out.push(seg("build-info", "build record".into(), b.offset, b.len));
6507    }
6508    for s in &f.root.sections {
6509        let Ok(a) = s.abs(h) else { continue };
6510        let kind: &'static str = match s.kind {
6511            crate::v06::kind::SCHEMA => "schema",
6512            crate::v06::kind::DICT_DATA => "dictionary",
6513            crate::v06::kind::INDEX_DATA => "tile",
6514            crate::v06::kind::QUAD_DATA => "named-graphs",
6515            crate::v06::kind::LABEL_DATA => "label-index",
6516            crate::v06::kind::TEXT_DATA => "text-index",
6517            crate::v06::kind::PYRAMID => "pyramid",
6518            _ => "directory",
6519        };
6520        out.push(seg(
6521            kind,
6522            crate::v06::kind::name(s.kind).to_string(),
6523            a.offset,
6524            a.len,
6525        ));
6526    }
6527    out.sort_by_key(|s| s.offset);
6528    out
6529}
6530
6531/// The deferred slots of a `0x06` open: pyramid, text index, label index and
6532/// schema, each faulted on first use.
6533#[allow(clippy::type_complexity)]
6534fn v06_slots(
6535    f: &std::sync::Arc<crate::v06::open::V06File>,
6536) -> (
6537    PyramidSlot,
6538    TextIndexSlot,
6539    SectionSlot<crate::label_index::LabelIndex>,
6540    SectionSlot<crate::schema_section::SchemaMeta>,
6541) {
6542    use crate::v06::{P_LABELS, P_PYRAMID, P_SCHEMA, P_TEXT};
6543    let pyramid = if f.h.has(P_PYRAMID) {
6544        let a = f.clone();
6545        PyramidSlot::Lazy {
6546            loader: Box::new(move || PyramidMeta::decode(&a.pyramid_bytes()?).ok()),
6547            cell: std::sync::OnceLock::new(),
6548        }
6549    } else {
6550        PyramidSlot::Resident(None)
6551    };
6552    let text_index = if f.h.has(P_TEXT) {
6553        let (a, b) = (f.clone(), f.clone());
6554        TextIndexSlot::Lazy {
6555            loader: Box::new(move || a.text_index()?.ok()),
6556            cell: std::sync::OnceLock::new(),
6557            token_table: Box::new(move || b.token_table_bytes()),
6558            token_table_cell: std::sync::OnceLock::new(),
6559        }
6560    } else {
6561        TextIndexSlot::Resident {
6562            index: None,
6563            token_table_len: None,
6564        }
6565    };
6566    let labels = if f.h.has(P_LABELS) {
6567        let a = f.clone();
6568        SectionSlot::Lazy {
6569            loader: Box::new(move || a.label_index()?.ok()),
6570            cell: std::sync::OnceLock::new(),
6571        }
6572    } else {
6573        SectionSlot::Resident(None)
6574    };
6575    let schema = if f.h.has(P_SCHEMA) {
6576        let a = f.clone();
6577        SectionSlot::Lazy {
6578            loader: Box::new(move || a.schema()),
6579            cell: std::sync::OnceLock::new(),
6580        }
6581    } else {
6582        SectionSlot::Resident(None)
6583    };
6584    (pyramid, text_index, labels, schema)
6585}
6586
6587/// The pyramid meta is large on real graphs (tens of MB) and SPARQL never reads
6588/// it, so a ranged open defers its fetch: it faults in only if `pyramid()` is
6589/// called (community / pyramid_tree / label-prefix queries).
6590fn ranged_pyramid_slot<R: RangeReader + Send + Sync + 'static>(
6591    reader: &std::sync::Arc<R>,
6592    header: &Header,
6593) -> PyramidSlot {
6594    if header.pyramid_meta_len == 0 {
6595        return PyramidSlot::Resident(None);
6596    }
6597    let pyr_reader = reader.clone();
6598    let pyr_off = header.pyramid_meta_offset;
6599    let pyr_len = header.pyramid_meta_len;
6600    PyramidSlot::Lazy {
6601        loader: Box::new(move || {
6602            let mb = pyr_reader.read_at(pyr_off, pyr_len).ok()?;
6603            PyramidMeta::decode(&mb).ok()
6604        }),
6605        cell: std::sync::OnceLock::new(),
6606    }
6607}
6608
6609/// The deferred TEXT_INDEX of a ranged open: a text search faults the token
6610/// table on first call (the leading varint then its compressed bytes), then
6611/// fetches individual posting lists by `(offset, len)` — never the whole
6612/// postings blob. A caller that never searches pays nothing.
6613fn ranged_text_index_slot<R: RangeReader + Send + Sync + 'static>(
6614    reader: &std::sync::Arc<R>,
6615    header: &Header,
6616) -> TextIndexSlot {
6617    let (ti_off, ti_len) = header.text_index_payload();
6618    if ti_len == 0 {
6619        return TextIndexSlot::Resident {
6620            index: None,
6621            token_table_len: None,
6622        };
6623    }
6624    let ti_reader = reader.clone();
6625    let probe_reader = reader.clone();
6626    let codec = header.block_codec;
6627    TextIndexSlot::Lazy {
6628        loader: Box::new(move || {
6629            // The section opens with `varint token_table_len`; read enough to
6630            // decode it (a uvarint is ≤ 10 bytes), then fetch the varint + the
6631            // compressed token table as one prefix range.
6632            let prefix_len = read_token_table_len(&*ti_reader, ti_off, ti_len)?;
6633            let prefix = ti_reader.read_at(ti_off, prefix_len).ok()?;
6634            let postings_base = crate::text_index::TextIndex::postings_base(&prefix)? as u64;
6635            let postings_abs = ti_off + postings_base;
6636            let pr = ti_reader.clone();
6637            let posting_loader =
6638                Box::new(move |off: u64, len: u64| pr.read_at(postings_abs + off, len).ok());
6639            crate::text_index::TextIndex::from_token_table(&prefix, codec, posting_loader).ok()
6640        }),
6641        cell: std::sync::OnceLock::new(),
6642        // The same leading varint the loader starts from, asked on its own so a
6643        // caller can *state* the cost of a first search without paying it.
6644        token_table: Box::new(move || read_token_table_len(&*probe_reader, ti_off, ti_len)),
6645        token_table_cell: std::sync::OnceLock::new(),
6646    }
6647}
6648
6649/// The `(label, subject term)` pairs of a legacy (pyramid-block) label search.
6650/// The hits' subjects are scattered over the dictionary, so their chunks are
6651/// faulted in one batch first (one round trip on a remote file, as
6652/// [`text_search_in`] does for full-text hits), then decoded in one
6653/// chunk-ordered pass. A subject that does not resolve drops its hit, as
6654/// before; the order is the search's.
6655fn resolve_label_hits(
6656    dict: &Dictionary,
6657    hits: Vec<&crate::meta::LabelEntry>,
6658) -> Vec<(String, String)> {
6659    let ids: Vec<u32> = hits.iter().map(|e| e.subject).collect();
6660    dict.prefetch_subject_terms(&ids);
6661    hits.iter()
6662        .zip(dict.subject_terms(&ids))
6663        .filter_map(|(e, iri)| Some((e.label.clone(), iri?)))
6664        .collect()
6665}
6666
6667/// Full-text search over `ti`, resolving the matched subject ids through `dict`.
6668/// The engine behind [`Rete::text_search`] and [`SearchView::text_search`].
6669fn text_search_in(
6670    ti: Option<&crate::text_index::TextIndex>,
6671    dict: &Dictionary,
6672    words: &[&str],
6673    prefix: Option<&str>,
6674    limit: usize,
6675) -> Vec<String> {
6676    let Some(ti) = ti else {
6677        return Vec::new();
6678    };
6679    // Each query word is tokenized exactly as at build time (so "Glucose"
6680    // matches the stored "glucose"); a word that splits into several tokens
6681    // requires all of them. AND across every required token + the prefix.
6682    let mut acc: Option<Vec<u32>> = None;
6683    if let Some(p) = prefix {
6684        acc = Some(ti.prefix(&p.to_lowercase()));
6685    }
6686    for w in words {
6687        for tok in crate::text_index::tokenize(w) {
6688            let posting = ti.lookup(&tok);
6689            acc = Some(match acc {
6690                Some(a) => intersect_sorted(&a, &posting),
6691                None => posting,
6692            });
6693            if acc.as_ref().is_some_and(|a| a.is_empty()) {
6694                return Vec::new();
6695            }
6696        }
6697    }
6698    let ids = acc.unwrap_or_default();
6699    let mut out = Vec::with_capacity(if limit > 0 {
6700        limit.min(ids.len())
6701    } else {
6702        ids.len()
6703    });
6704    // The hits' subject IRIs are scattered over many dictionary chunks: fault
6705    // the ones a page can return in one batch (on a remote file, one round
6706    // trip instead of one per chunk). Only warms the cache.
6707    let page = if limit > 0 {
6708        limit.min(ids.len())
6709    } else {
6710        ids.len()
6711    };
6712    dict.prefetch_subject_terms(&ids[..page]);
6713    for id in ids {
6714        if let Some(iri) = dict.subject_term(id) {
6715            out.push(iri);
6716            if limit > 0 && out.len() >= limit {
6717                break;
6718            }
6719        }
6720    }
6721    out
6722}
6723
6724/// A `.rete` opened for **search only**: the dictionary's chunk directories, and
6725/// the TEXT_INDEX / pyramid deferred behind their lazy slots. Nothing else.
6726///
6727/// What makes this cheap is skipping dictionary sections 2 and 3 (object-only
6728/// and predicates) — NOT skipping the index. Measured on the published
6729/// `epfl-infoscience.rete` (1.64 GB), a full [`Rete::open_ranged_lazy`] reads
6730/// 536,947,344 B, of which the six permutation tile directories are **49,940 B
6731/// in 40 reads** and the dictionary is 536,896,380 B in 36 reads: the
6732/// object-only chunk directory alone is 234,400,728 B, because that file stores
6733/// every chunk's first term verbatim. Files built since #198 key it by the
6734/// shortest separator (~600 KB on this graph); the published one still does
6735/// not, and only a rebuild changes that. This view opens sections 0 and 1
6736/// only — both search modes return subject IRIs — and costs **21,554 B in 9
6737/// reads** on the same file. It also skips the index container and defers the
6738/// TEXT_INDEX and pyramid, so it pays only for what it reads: the token table
6739/// on the first search, then one range per posting list and per dictionary
6740/// chunk holding a matched subject. This is what backs `rete search-url`.
6741///
6742/// On a file with a LABEL_INDEX section (kind 8) the dictionary is not opened
6743/// at all until a full-text search needs it: the section stores each subject's
6744/// term, so a prefix search costs the header, the section head, one directory
6745/// node per level and one data block.
6746pub struct SearchView {
6747    header: Header,
6748    dict: SectionSlot<Dictionary>,
6749    pyramid: PyramidSlot,
6750    text_index: TextIndexSlot,
6751    labels: SectionSlot<crate::label_index::LabelIndex>,
6752}
6753
6754impl SearchView {
6755    /// Open for search over any [`RangeReader`] (HTTP or local): header +
6756    /// dictionary chunk directories only (none at all on a file with a
6757    /// LABEL_INDEX, until a text search needs them), with the text index,
6758    /// label index and pyramid faulting in on first use.
6759    pub fn open_ranged<R: RangeReader + Send + Sync + 'static>(
6760        reader: R,
6761    ) -> Result<Self, FileError> {
6762        Self::open_ranged_with(reader, crate::v06::FIRST_READ)
6763    }
6764
6765    /// [`open_ranged`](Self::open_ranged) with an explicit first-read size
6766    /// (see [`Rete::open_ranged_lazy_with`]).
6767    pub fn open_ranged_with<R: RangeReader + Send + Sync + 'static>(
6768        reader: R,
6769        first_read: u64,
6770    ) -> Result<Self, FileError> {
6771        let first_len = first_read.max(HEADER_LEN as u64).min(reader.len());
6772        let first: std::sync::Arc<[u8]> = std::sync::Arc::from(reader.read_at(0, first_len)?);
6773        let reader = crate::reader::PrefixReader::new(reader, first.clone());
6774        if crate::v06::open::is_v06(&first) {
6775            let r: std::sync::Arc<dyn RangeReader + Send + Sync> = std::sync::Arc::new(reader);
6776            let f = std::sync::Arc::new(crate::v06::open::V06File::open(r, &first)?);
6777            let header = crate::v06::logical_header(&f.h);
6778            let (pyramid, text_index, labels, _) = v06_slots(&f);
6779            let want = [true, true, false, false];
6780            let dict = if f.h.has(crate::v06::P_LABELS) {
6781                let a = f.clone();
6782                SectionSlot::Lazy {
6783                    loader: Box::new(move || a.dictionary(want).ok()),
6784                    cell: std::sync::OnceLock::new(),
6785                }
6786            } else {
6787                SectionSlot::Resident(Some(f.dictionary(want)?))
6788            };
6789            return Ok(Self {
6790                header,
6791                dict,
6792                pyramid,
6793                text_index,
6794                labels,
6795            });
6796        }
6797        let head = reader.read_at(0, HEADER_LEN as u64)?;
6798        let header = Header::from_bytes(&head)?;
6799        let reader = std::sync::Arc::new(reader);
6800        // Sections 0 (shared) and 1 (subject-only) are the only ones a search
6801        // resolves — both modes return subject IRIs. Skipping 2 and 3 is what
6802        // makes this open cheap where a full one is not.
6803        let dict = if header.label_index_section().is_some() {
6804            let (r, h) = (reader.clone(), header.clone());
6805            SectionSlot::Lazy {
6806                loader: Box::new(move || {
6807                    ranged_chunked_dictionary(&r, &h, [true, true, false, false]).ok()
6808                }),
6809                cell: std::sync::OnceLock::new(),
6810            }
6811        } else {
6812            SectionSlot::Resident(Some(ranged_chunked_dictionary(
6813                &reader,
6814                &header,
6815                [true, true, false, false],
6816            )?))
6817        };
6818        let pyramid = ranged_pyramid_slot(&reader, &header);
6819        let text_index = ranged_text_index_slot(&reader, &header);
6820        let labels = ranged_label_slot(&reader, &header);
6821        Ok(Self {
6822            header,
6823            dict,
6824            pyramid,
6825            text_index,
6826            labels,
6827        })
6828    }
6829
6830    /// Whether the file can answer a prefix search: it carries a LABEL_INDEX
6831    /// section or a pyramid (whose meta may hold the legacy label block). A
6832    /// header read, no fault.
6833    pub fn has_labels(&self) -> bool {
6834        self.header.label_index_section().is_some() || self.header.pyramid_meta_len > 0
6835    }
6836
6837    /// Whether the file carries a LABEL_INDEX section (kind 8).
6838    pub fn has_label_index(&self) -> bool {
6839        self.header.label_index_section().is_some()
6840    }
6841
6842    /// The file header (already fetched by [`open_ranged`](Self::open_ranged)).
6843    pub fn header(&self) -> &Header {
6844        &self.header
6845    }
6846
6847    /// Whether the file carries a TEXT_INDEX section — a header read, no fault.
6848    pub fn has_text_index(&self) -> bool {
6849        self.header.text_index_len > 0
6850    }
6851
6852    /// Whether the file carries a pyramid (where the label index lives) — a
6853    /// header read, no fault.
6854    pub fn has_pyramid(&self) -> bool {
6855        self.header.pyramid_meta_len > 0
6856    }
6857
6858    /// Full-text search — see [`Rete::text_search`] for the semantics. Faults
6859    /// the token table on the first call, then only the queried postings.
6860    pub fn text_search(&self, words: &[&str], prefix: Option<&str>, limit: usize) -> Vec<String> {
6861        let Some(dict) = self.dict.get() else {
6862            return Vec::new();
6863        };
6864        text_search_in(self.text_index.index(), dict, words, prefix, limit)
6865    }
6866
6867    /// What that first search will cost — see
6868    /// [`Rete::text_index_token_table_len`]. One ≤10-byte range read, so a
6869    /// caller can quote the figure before deciding to search.
6870    pub fn text_index_token_table_len(&self) -> Option<u64> {
6871        self.text_index.token_table_len()
6872    }
6873
6874    /// Label-prefix search — see [`Rete::prefix_search`]. Reads the
6875    /// LABEL_INDEX section when the file has one; otherwise faults the pyramid
6876    /// (which carries the legacy label block), the pricier path.
6877    pub fn prefix_search(&self, prefix: &str, limit: usize) -> Vec<(String, String)> {
6878        if let Some(li) = self.labels.get() {
6879            return li.prefix_search(prefix, limit);
6880        }
6881        let pyr = match &self.pyramid {
6882            PyramidSlot::Resident(p) => p.as_ref(),
6883            PyramidSlot::Lazy { loader, cell } => cell.get_or_init(loader).as_ref(),
6884        };
6885        let (Some(pyr), Some(dict)) = (pyr, self.dict.get()) else {
6886            return Vec::new();
6887        };
6888        resolve_label_hits(dict, pyr.prefix_search(prefix, limit))
6889    }
6890}
6891
6892/// A lightweight, overview-only view of a file: the pyramid summary graph plus
6893/// just enough dictionary to label predicates. Fetched via ranges *without*
6894/// touching the (large) triple index — the "load the coarse graph first" path
6895/// from SPEC.md §7.2.
6896#[must_use]
6897pub struct SummaryView {
6898    pub round: u32,
6899    pub summary: Vec<SuperEdge>,
6900    /// The shipped `subClassOf` hierarchy (v2 schema pyramid; empty on v1 files).
6901    pub class_hierarchy: Vec<ClassNode>,
6902    /// Per-level type rollups — the leveled legend, read index-free.
6903    pub level_rollups: Vec<LevelRollup>,
6904    /// Per-level lateral class-relation graph (the non-`is-a` connections).
6905    pub level_links: Vec<LevelLinks>,
6906    /// Per-community descriptors (Phase 4 progressive refinement; may be empty).
6907    pub descriptors: Vec<CommunityDescriptor>,
6908    /// `subClassOf` cycles (v2.1; empty on older files).
6909    pub subclass_cycles: Vec<Vec<String>>,
6910    /// `owl:disjointWith` class pairs (v2.1; empty on older files).
6911    pub disjoint_pairs: Vec<(String, String)>,
6912    /// `owl:equivalentClass` class pairs (v2.1; empty on older files).
6913    pub equivalent_pairs: Vec<(String, String)>,
6914    dict: Dictionary,
6915}
6916
6917impl SummaryView {
6918    /// Read header → dictionary → pyramid-meta only (skips the index container).
6919    pub fn open_ranged<R: RangeReader>(reader: &R) -> Result<Option<Self>, FileError> {
6920        let head = reader.read_at(0, (HEADER_LEN as u64).min(reader.len()))?;
6921        let header = Header::from_bytes(&head)?;
6922        if header.pyramid_meta_len == 0 {
6923            return Ok(None);
6924        }
6925        if header.version == crate::v06::VERSION {
6926            return Self::open_ranged_v06(reader, &header, head).map(Some);
6927        }
6928
6929        let dict_bytes = reader.read_at(header.dictionary_offset, header.dictionary_len)?;
6930        let dict = decode_dictionary_container(&dict_bytes, header.dict_codec)?;
6931
6932        let mb = reader.read_at(header.pyramid_meta_offset, header.pyramid_meta_len)?;
6933        let mut meta =
6934            PyramidMeta::decode(&mb).map_err(|_| FileError::Container("malformed pyramid meta"))?;
6935        // A file whose schema moved to its own section (kind 9) keeps only the
6936        // community material here: take the schema fields from that section.
6937        if let Some(s) = read_schema_with_header(reader, &header) {
6938            meta.class_hierarchy = s.class_hierarchy;
6939            meta.level_rollups = s.level_rollups;
6940            meta.level_links = s.level_links;
6941            meta.subclass_cycles = s.subclass_cycles;
6942            meta.disjoint_pairs = s.disjoint_pairs;
6943            meta.equivalent_pairs = s.equivalent_pairs;
6944        }
6945
6946        Ok(Some(SummaryView {
6947            round: meta.round,
6948            summary: meta.summary,
6949            class_hierarchy: meta.class_hierarchy,
6950            level_rollups: meta.level_rollups,
6951            level_links: meta.level_links,
6952            descriptors: meta.descriptors,
6953            subclass_cycles: meta.subclass_cycles,
6954            disjoint_pairs: meta.disjoint_pairs,
6955            equivalent_pairs: meta.equivalent_pairs,
6956            dict,
6957        }))
6958    }
6959
6960    /// [`open_ranged`](Self::open_ranged) for a `0x06` file: the pyramid
6961    /// section, the schema, and the dictionary (its leaves and chunks read
6962    /// whole, as the `0x05` path reads the whole dictionary container).
6963    fn open_ranged_v06<R: RangeReader>(
6964        reader: &R,
6965        header: &Header,
6966        head: Vec<u8>,
6967    ) -> Result<Self, FileError> {
6968        use crate::v06::{kind, open};
6969        // One first read, then every range the view needs past it — the
6970        // pyramid, the schema, the dictionary's leaves and chunks — in one
6971        // batch.
6972        let mut first = head;
6973        let want_first = crate::v06::FIRST_READ.min(reader.len());
6974        if want_first > first.len() as u64 {
6975            let at = first.len() as u64;
6976            first.extend_from_slice(&reader.read_at(at, want_first - at)?);
6977        }
6978        let pre = open::Prefixed {
6979            first: &first,
6980            r: reader,
6981        };
6982        let (h, root) = open::read_front(&pre)?;
6983        let want: Vec<ByteRange> = [
6984            kind::PYRAMID,
6985            kind::SCHEMA,
6986            kind::DICT_LEAVES,
6987            kind::DICT_DATA,
6988        ]
6989        .iter()
6990        .filter_map(|&k| root.section(k).and_then(|s| s.abs(&h).ok()))
6991        .collect();
6992        let sparse = open::SparseReader::fetch(&pre, first.clone(), &want)?;
6993        let mb = open::read_section(&sparse, &h, &root, kind::PYRAMID)?
6994            .ok_or(FileError::Container("0x06 file has no pyramid"))?;
6995        let mut meta =
6996            PyramidMeta::decode(&mb).map_err(|_| FileError::Container("malformed pyramid meta"))?;
6997        if let Some(s) = read_schema_with_header(&sparse, header) {
6998            meta.class_hierarchy = s.class_hierarchy;
6999            meta.level_rollups = s.level_rollups;
7000            meta.level_links = s.level_links;
7001            meta.subclass_cycles = s.subclass_cycles;
7002            meta.disjoint_pairs = s.disjoint_pairs;
7003            meta.equivalent_pairs = s.equivalent_pairs;
7004        }
7005        let f = open::V06File::open(std::sync::Arc::new(sparse), &first)?;
7006        let dict = f.dictionary([true; 4])?;
7007        Ok(SummaryView {
7008            round: meta.round,
7009            summary: meta.summary,
7010            class_hierarchy: meta.class_hierarchy,
7011            level_rollups: meta.level_rollups,
7012            level_links: meta.level_links,
7013            descriptors: meta.descriptors,
7014            subclass_cycles: meta.subclass_cycles,
7015            disjoint_pairs: meta.disjoint_pairs,
7016            equivalent_pairs: meta.equivalent_pairs,
7017            dict,
7018        })
7019    }
7020
7021    /// A schema-only view of a file **without a pyramid**, from its SCHEMA
7022    /// section (kind 9): the hierarchy, levels and axioms, with an empty
7023    /// community summary. Reads the header and the SCHEMA section only — the
7024    /// section carries its own class and predicate strings, so no dictionary.
7025    /// `Ok(None)` when the file has no SCHEMA section. Callers that need the
7026    /// summary (`cost`, federation source selection, progressive queries) use
7027    /// [`open_ranged`](Self::open_ranged), which keeps requiring a pyramid.
7028    pub fn open_schema_ranged<R: RangeReader>(reader: &R) -> Result<Option<Self>, FileError> {
7029        let head = reader.read_at(0, HEADER_LEN as u64)?;
7030        let header = Header::from_bytes(&head)?;
7031        let Some(s) = read_schema_with_header(reader, &header) else {
7032            return Ok(None);
7033        };
7034        let dict = crate::DictionaryBuilder::new().build();
7035        Ok(Some(SummaryView {
7036            round: 0,
7037            summary: Vec::new(),
7038            class_hierarchy: s.class_hierarchy,
7039            level_rollups: s.level_rollups,
7040            level_links: s.level_links,
7041            descriptors: Vec::new(),
7042            subclass_cycles: s.subclass_cycles,
7043            disjoint_pairs: s.disjoint_pairs,
7044            equivalent_pairs: s.equivalent_pairs,
7045            dict,
7046        }))
7047    }
7048
7049    /// Number of semantic-zoom levels in the schema pyramid (0 if none shipped).
7050    pub fn level_count(&self) -> usize {
7051        self.level_rollups.len()
7052    }
7053
7054    /// The type rollup at semantic level `k` (0 = coarsest/most abstract), or
7055    /// `None` if `k` is out of range. Index-free — answered from the pyramid-meta.
7056    pub fn level_rollup(&self, k: usize) -> Option<&LevelRollup> {
7057        self.level_rollups.get(k)
7058    }
7059
7060    /// Resolve a predicate ID in the summary to its term.
7061    pub fn predicate_term(&self, id: u32) -> Option<String> {
7062        self.dict.predicate_term(id)
7063    }
7064
7065    /// Exact number of triples using `predicate`, summed from the summary's
7066    /// superedge counts — answered without ever reading the triple index.
7067    pub fn predicate_total(&self, predicate: &str) -> u32 {
7068        match self.dict.predicate_id(predicate) {
7069            Some(pid) => self
7070                .summary
7071                .iter()
7072                .filter(|e| e.predicate == pid)
7073                .map(|e| e.count)
7074                .sum(),
7075            None => 0,
7076        }
7077    }
7078
7079    /// All predicates with their exact triple totals, descending by count.
7080    pub fn predicate_totals(&self) -> Vec<(String, u32)> {
7081        let mut by_pred: std::collections::BTreeMap<u32, u32> = std::collections::BTreeMap::new();
7082        for e in &self.summary {
7083            *by_pred.entry(e.predicate).or_default() += e.count;
7084        }
7085        let mut out: Vec<(String, u32)> = by_pred
7086            .into_iter()
7087            .filter_map(|(pid, c)| self.dict.predicate_term(pid).map(|t| (t, c)))
7088            .collect();
7089        out.sort_by(|a, b| b.1.cmp(&a.1).then_with(|| a.0.cmp(&b.0)));
7090        out
7091    }
7092
7093    /// Number of communities the summary spans (distinct supernode endpoints).
7094    pub fn community_count(&self) -> usize {
7095        let mut comms = std::collections::BTreeSet::new();
7096        for e in &self.summary {
7097            comms.insert(e.s_comm);
7098            comms.insert(e.o_comm);
7099        }
7100        comms.len()
7101    }
7102
7103    /// **Index-free T-Box coherence (Tier-0).** Detect schema-level incoherent
7104    /// points purely from the shipped schema pyramid — no triple index, no
7105    /// instance data, O(ontology) regardless of graph size:
7106    /// - `subclass-cycle`: a set of classes that are mutually `rdfs:subClassOf`.
7107    /// - `unsatisfiable-class`: a class whose ancestor closure (over all parents,
7108    ///   folded through `owl:equivalentClass`) contains both ends of an
7109    ///   `owl:disjointWith` pair, so no individual can ever be one.
7110    ///
7111    /// Soundness is bounded by what the pyramid ships: the `subClassOf` hierarchy
7112    /// is capped (`MAX_HIERARCHY` in `schema_pyramid`), so on a very large ontology
7113    /// a pruned ancestor can hide an unsatisfiable class (a false *coherent*, never
7114    /// a false *incoherent*). Instance-level clashes (a node typed into disjoint
7115    /// classes, functional-property clashes) are NOT visible here — they need the
7116    /// A-Box (Tier-1/Tier-2 `reason`).
7117    pub fn tbox_coherence(&self) -> Vec<crate::reason::Inconsistency> {
7118        schema_coherence(
7119            &self.class_hierarchy,
7120            &self.subclass_cycles,
7121            &self.disjoint_pairs,
7122            &self.equivalent_pairs,
7123        )
7124    }
7125
7126    /// True when [`tbox_coherence`](Self::tbox_coherence) finds no schema-level
7127    /// incoherent point.
7128    pub fn tbox_is_coherent(&self) -> bool {
7129        self.tbox_coherence().is_empty()
7130    }
7131}
7132
7133/// Compute T-Box coherence points from the schema-pyramid fields alone — no
7134/// dictionary, no index, no instance data. Shared by [`SummaryView::tbox_coherence`]
7135/// and the dictionary-free [`read_schema_coherence_ranged`]. Emits `subclass-cycle`
7136/// and `unsatisfiable-class` (a class whose ancestor closure — over all parents,
7137/// folded through `owl:equivalentClass` — contains both ends of a disjoint pair).
7138pub fn schema_coherence(
7139    class_hierarchy: &[ClassNode],
7140    subclass_cycles: &[Vec<String>],
7141    disjoint_pairs: &[(String, String)],
7142    equivalent_pairs: &[(String, String)],
7143) -> Vec<crate::reason::Inconsistency> {
7144    use crate::reason::Inconsistency;
7145    use std::collections::{BTreeMap, BTreeSet, VecDeque};
7146    const MAX_REACH: usize = 100_000;
7147
7148    let mut out: Vec<Inconsistency> = Vec::new();
7149
7150    for cyc in subclass_cycles {
7151        let detail = if cyc.len() == 1 {
7152            format!("{} is rdfs:subClassOf itself (a cycle)", cyc[0])
7153        } else {
7154            format!(
7155                "classes {{{}}} are mutually rdfs:subClassOf (a cycle)",
7156                cyc.join(", ")
7157            )
7158        };
7159        out.push(Inconsistency {
7160            kind: "subclass-cycle",
7161            detail,
7162        });
7163    }
7164
7165    if !disjoint_pairs.is_empty() {
7166        // Upward adjacency: subClassOf parents + bidirectional equivalence.
7167        let mut adj: BTreeMap<&str, Vec<&str>> = BTreeMap::new();
7168        for n in class_hierarchy {
7169            let e = adj.entry(n.class.as_str()).or_default();
7170            for p in &n.parents {
7171                e.push(p.as_str());
7172            }
7173        }
7174        for (a, b) in equivalent_pairs {
7175            adj.entry(a.as_str()).or_default().push(b.as_str());
7176            adj.entry(b.as_str()).or_default().push(a.as_str());
7177        }
7178
7179        // Candidate focus classes: every class named anywhere in the schema.
7180        let mut focuses: BTreeSet<&str> =
7181            class_hierarchy.iter().map(|n| n.class.as_str()).collect();
7182        for (a, b) in disjoint_pairs.iter().chain(equivalent_pairs) {
7183            focuses.insert(a.as_str());
7184            focuses.insert(b.as_str());
7185        }
7186
7187        let mut seen: BTreeSet<&str> = BTreeSet::new();
7188        for &c in &focuses {
7189            // reach(c) = {c} ∪ ancestors (capped BFS over `adj`).
7190            let mut reach: BTreeSet<&str> = BTreeSet::new();
7191            let mut q: VecDeque<&str> = VecDeque::new();
7192            reach.insert(c);
7193            q.push_back(c);
7194            while let Some(x) = q.pop_front() {
7195                if reach.len() > MAX_REACH {
7196                    break;
7197                }
7198                if let Some(ns) = adj.get(x) {
7199                    for &p in ns {
7200                        if reach.insert(p) {
7201                            q.push_back(p);
7202                        }
7203                    }
7204                }
7205            }
7206            for (x, y) in disjoint_pairs {
7207                if reach.contains(x.as_str()) && reach.contains(y.as_str()) && seen.insert(c) {
7208                    out.push(Inconsistency {
7209                        kind: "unsatisfiable-class",
7210                        detail: format!(
7211                            "{c} is a subclass of both {x} and {y}, which are \
7212                             owl:disjointWith — no individual can be a {c}"
7213                        ),
7214                    });
7215                    break;
7216                }
7217            }
7218        }
7219    }
7220
7221    out.sort_by(|a, b| (a.kind, &a.detail).cmp(&(b.kind, &b.detail)));
7222    out
7223}
7224
7225/// **Dictionary-free Tier-0 coherence read.** Fetch only the header and the
7226/// pyramid-meta range (2 small range reads) and run [`schema_coherence`] over the
7227/// schema pyramid — never touching the **dictionary** (which a literal-heavy file
7228/// makes large) or the triple index. `Ok(None)` if the file ships no pyramid.
7229///
7230/// This is what makes the Tier-0 check cheap on big graphs: the schema pyramid
7231/// carries its own class-string table, so coherence needs none of the dictionary.
7232pub fn read_schema_coherence_ranged<R: RangeReader>(
7233    reader: &R,
7234) -> Result<Option<Vec<crate::reason::Inconsistency>>, FileError> {
7235    let head = reader.read_at(0, HEADER_LEN as u64)?;
7236    let header = Header::from_bytes(&head)?;
7237    // The SCHEMA section (kind 9) first: one range, and no pyramid needed.
7238    if let Some(s) = read_schema_with_header(reader, &header) {
7239        return Ok(Some(schema_coherence(
7240            &s.class_hierarchy,
7241            &s.subclass_cycles,
7242            &s.disjoint_pairs,
7243            &s.equivalent_pairs,
7244        )));
7245    }
7246    if header.pyramid_meta_len == 0 {
7247        return Ok(None);
7248    }
7249    // Fast path: the header records the trailing schema block's length, so read ONLY
7250    // that block (at the end of pyramid-meta) — never the community summary, the
7251    // dictionary, or the index. This is what makes Tier-0 flat at any graph size.
7252    if header.schema_meta_len > 0 && (header.schema_meta_len as u64) <= header.pyramid_meta_len {
7253        let off =
7254            header.pyramid_meta_offset + header.pyramid_meta_len - header.schema_meta_len as u64;
7255        let block = reader.read_at(off, header.schema_meta_len as u64)?;
7256        let (hierarchy, cycles, disjoint, equivalent) = crate::meta::decode_schema_block(&block)
7257            .map_err(|_| FileError::Container("malformed schema block"))?;
7258        return Ok(Some(schema_coherence(
7259            &hierarchy,
7260            &cycles,
7261            &disjoint,
7262            &equivalent,
7263        )));
7264    }
7265    // Fallback (pre-v0.2.1 files with no header field): decode the whole pyramid-meta.
7266    let mb = if header.version == crate::v06::VERSION {
7267        crate::v06::open::read_pyramid(reader)?.unwrap_or_default()
7268    } else {
7269        reader.read_at(header.pyramid_meta_offset, header.pyramid_meta_len)?
7270    };
7271    let meta =
7272        PyramidMeta::decode(&mb).map_err(|_| FileError::Container("malformed pyramid meta"))?;
7273    Ok(Some(schema_coherence(
7274        &meta.class_hierarchy,
7275        &meta.subclass_cycles,
7276        &meta.disjoint_pairs,
7277        &meta.equivalent_pairs,
7278    )))
7279}
7280
7281/// The **schema summary** (per-class histogram + class relations at the finest
7282/// level) read over a [`RangeReader`] from the schema pyramid alone — the
7283/// index-free, range-readable source for a Schema view of a remote graph. Returns
7284/// `(classes, relations)` with `classes = [(class_iri, count)]` and `relations =
7285/// [(s_class, predicate, o_class, count)]`; `None` when the file has no schema
7286/// pyramid. Like [`read_schema_coherence_ranged`], it reads only the trailing
7287/// schema block, so it stays flat at any graph size.
7288#[allow(clippy::type_complexity)]
7289pub fn read_schema_summary_ranged<R: RangeReader>(
7290    reader: &R,
7291) -> Result<Option<(Vec<(String, u64)>, Vec<(String, String, String, u64)>)>, FileError> {
7292    let head = reader.read_at(0, HEADER_LEN as u64)?;
7293    let header = Header::from_bytes(&head)?;
7294    // The SCHEMA section (kind 9) first: one range, and no pyramid needed.
7295    let meta = if let Some(s) = read_schema_with_header(reader, &header) {
7296        PyramidMeta::new(0, Vec::new(), &[]).with_schema(
7297            Vec::new(),
7298            s.level_rollups,
7299            s.level_links,
7300            Vec::new(),
7301            Vec::new(),
7302            Vec::new(),
7303            Vec::new(),
7304        )
7305    } else {
7306        if header.pyramid_meta_len == 0 {
7307            return Ok(None);
7308        }
7309        if header.schema_meta_len > 0 && (header.schema_meta_len as u64) <= header.pyramid_meta_len
7310        {
7311            let off = header.pyramid_meta_offset + header.pyramid_meta_len
7312                - header.schema_meta_len as u64;
7313            let block = reader.read_at(off, header.schema_meta_len as u64)?;
7314            let summary = crate::meta::decode_schema_block_summary(&block)
7315                .map_err(|_| FileError::Container("malformed schema block"))?;
7316            return Ok(Some(summary));
7317        }
7318        // Fallback (pre-v0.2.1 files): decode the whole pyramid-meta, pull finest levels.
7319        let mb = if header.version == crate::v06::VERSION {
7320            crate::v06::open::read_pyramid(reader)?.unwrap_or_default()
7321        } else {
7322            reader.read_at(header.pyramid_meta_offset, header.pyramid_meta_len)?
7323        };
7324        PyramidMeta::decode(&mb).map_err(|_| FileError::Container("malformed pyramid meta"))?
7325    };
7326    if meta.level_rollups.is_empty() && meta.level_links.is_empty() {
7327        return Ok(None);
7328    }
7329    let classes = meta
7330        .level_rollups
7331        .iter()
7332        .max_by_key(|r| r.depth)
7333        .map(|r| r.classes.clone())
7334        .unwrap_or_default();
7335    let relations = meta
7336        .level_links
7337        .iter()
7338        .max_by_key(|l| l.depth)
7339        .map(|l| {
7340            l.links
7341                .iter()
7342                .map(|c| {
7343                    (
7344                        c.s_class.clone(),
7345                        c.predicate.clone(),
7346                        c.o_class.clone(),
7347                        c.count,
7348                    )
7349                })
7350                .collect()
7351        })
7352        .unwrap_or_default();
7353    Ok(Some((classes, relations)))
7354}
7355
7356/// Every dictionary section's chunk **routing keys**, exactly as stored, from a
7357/// finished file image. Test-only: it exists so a test in another module of
7358/// this crate (`extbuild`) can assert on the directory the writers produce
7359/// without any of this becoming public surface.
7360#[cfg(test)]
7361pub(crate) fn dict_chunk_keys_for_test(image: &[u8]) -> Vec<Vec<Vec<u8>>> {
7362    if crate::v06::open::is_v06(image) {
7363        // 0x06: the routing keys sit in the leaf directories of the dictionary.
7364        use crate::v06::{decode_chunk_leaf, decode_dict_root, kind, open, tag};
7365        let r = crate::reader::SliceReader::new(image);
7366        let (h, root) = open::read_front(&r).unwrap();
7367        let roots = decode_dict_root(root.unit(tag::DICT).unwrap()).unwrap();
7368        let leaves = open::read_section(&r, &h, &root, kind::DICT_LEAVES)
7369            .unwrap()
7370            .unwrap();
7371        return roots
7372            .iter()
7373            .map(|s| {
7374                s.leaves
7375                    .iter()
7376                    .flat_map(|(l, _, _, _)| {
7377                        let stored = &leaves[l.off as usize..(l.off + l.len) as usize];
7378                        let raw = decompress(h.internal_codec, stored).unwrap();
7379                        decode_chunk_leaf(&raw, l.n).unwrap()
7380                    })
7381                    .map(|c| c.key)
7382                    .collect()
7383            })
7384            .collect();
7385    }
7386    let header = Header::from_bytes(&image[..HEADER_LEN]).unwrap();
7387    let s = header.dictionary_offset as usize;
7388    let e = s + header.dictionary_len as usize;
7389    decode_container(&image[s..e], CODEC_NONE)
7390        .unwrap()
7391        .iter()
7392        .map(|payload| {
7393            let (_meta, entries) = parse_chunked_dict_dir(payload, payload.len() as u64).unwrap();
7394            entries.into_iter().map(|c| c.key).collect()
7395        })
7396        .collect()
7397}
7398
7399#[cfg(test)]
7400mod tests {
7401    use super::*;
7402    use super::{
7403        write_dataset_v05 as write_dataset,
7404        write_dataset_with_metadata_v05 as write_dataset_with_metadata,
7405        write_file_v05 as write_file,
7406    };
7407    use crate::dictionary::DictionaryBuilder;
7408    use crate::index::GraphIndexBuilder;
7409
7410    /// The pure-Rust decode path (wasm, `--no-default-features`) reads a zstd
7411    /// frame back byte for byte. Guards the `ruzstd` API across upgrades.
7412    #[cfg(not(feature = "compression"))]
7413    #[test]
7414    fn ruzstd_decodes_a_zstd_frame() {
7415        let data: Vec<u8> = (0..50_000u32)
7416            .flat_map(|i| (i % 251).to_le_bytes())
7417            .collect();
7418        let frame = ruzstd::encoding::compress_to_vec(
7419            &data[..],
7420            ruzstd::encoding::CompressionLevel::Fastest,
7421        );
7422        assert!(frame.len() < data.len());
7423        assert_eq!(decompress(CODEC_ZSTD, &frame).unwrap(), data);
7424    }
7425
7426    /// Bounded decompression: a frame whose header declares more than the
7427    /// bound is refused before decoding; one that states no size is decoded at
7428    /// most one byte past the bound; one that lies about its size fails; and
7429    /// the size a header states is read without decoding.
7430    #[cfg(feature = "compression")]
7431    #[test]
7432    fn decompression_is_bounded() {
7433        let data: Vec<u8> = (0..100_000u32).map(|i| (i % 251) as u8).collect();
7434        let sized = zstd::bulk::compress(&data, 3).unwrap();
7435        assert_eq!(zstd_frame_content_size(&sized), Some(100_000));
7436        assert_eq!(block_raw_len(CODEC_ZSTD, &sized, 1 << 20).unwrap(), 100_000);
7437        assert_eq!(
7438            decompress_bounded(CODEC_ZSTD, &sized, 100_000).unwrap(),
7439            data
7440        );
7441        assert!(decompress_bounded(CODEC_ZSTD, &sized, 99_999).is_err());
7442        // A streamed frame states no size: bounded by decoding.
7443        let stream = zstd::encode_all(&data[..], 3).unwrap();
7444        assert_eq!(zstd_frame_content_size(&stream), None);
7445        assert_eq!(
7446            block_raw_len(CODEC_ZSTD, &stream, 1 << 20).unwrap(),
7447            100_000
7448        );
7449        assert_eq!(
7450            decompress_bounded(CODEC_ZSTD, &stream, 100_000).unwrap(),
7451            data
7452        );
7453        assert!(decompress_bounded(CODEC_ZSTD, &stream, 50_000).is_err());
7454        // A header that under-states its content fails instead of growing.
7455        let mut liar = sized.clone();
7456        let fhd = liar[4];
7457        let fcs_at = 5 + usize::from(fhd & 0x20 == 0);
7458        liar[fcs_at] = liar[fcs_at].wrapping_sub(1);
7459        assert!(decompress_bounded(CODEC_ZSTD, &liar, 200_000).is_err());
7460        // Uncompressed blocks: the length is the size.
7461        assert!(decompress_bounded(CODEC_NONE, &data, 99_999).is_err());
7462        assert_eq!(
7463            decompress_bounded(CODEC_NONE, &data, 100_000).unwrap(),
7464            data
7465        );
7466        assert_eq!(zstd_frame_content_size(b"RETE"), None);
7467    }
7468
7469    /// A three-permutation file must survive every opener — resident, ranged,
7470    /// ranged-lazy — and answer every pattern shape with the rows the
7471    /// six-permutation twin gives, including through a named graph's own
7472    /// container. The mask is a file-level fact, so getting it to the named
7473    /// graphs' decoder is a separate thing that can be forgotten.
7474    #[test]
7475    fn three_permutation_file_round_trips_every_open_path() {
7476        use crate::index::PermSet;
7477        use crate::reader::SliceReader;
7478
7479        let build = |perms: PermSet| {
7480            let mut db = crate::DictionaryBuilder::new();
7481            let mut triples = Vec::new();
7482            for i in 0..60u32 {
7483                let (s, p, o) = (
7484                    format!("<http://ex/s{}>", i % 11),
7485                    format!("<http://ex/p{}>", i % 3),
7486                    format!("<http://ex/o{}>", i % 7),
7487                );
7488                db.observe(&s, &p, &o);
7489                triples.push((s, p, o));
7490            }
7491            let dict = db.build();
7492            let ids: Vec<(u32, u32, u32)> = triples
7493                .iter()
7494                .map(|(s, p, o)| dict.encode(s, p, o).unwrap())
7495                .collect();
7496            let def = GraphIndexBuilder::from_triples(ids.clone())
7497                .with_tile_budget(96)
7498                .with_perms(perms)
7499                .build();
7500            let named = GraphIndexBuilder::from_triples(ids)
7501                .with_tile_budget(96)
7502                .with_perms(perms)
7503                .build();
7504            write_dataset(
7505                &dict,
7506                &def,
7507                &[("<http://ex/g>".to_string(), named)],
7508                true,
7509                &[],
7510                0,
7511            )
7512        };
7513
7514        let six = build(PermSet::ALL);
7515        let three = build(PermSet::CORE);
7516        assert!(three.len() < six.len());
7517        assert_eq!(Header::from_bytes(&three).unwrap().perms, PermSet::CORE);
7518        assert_eq!(Header::from_bytes(&six).unwrap().perms, PermSet::ALL);
7519
7520        let shapes: Vec<(Option<&str>, Option<&str>, Option<&str>)> = vec![
7521            (None, None, None),
7522            (Some("<http://ex/s3>"), None, None),
7523            (None, Some("<http://ex/p1>"), None),
7524            (None, None, Some("<http://ex/o5>")),
7525            (Some("<http://ex/s3>"), Some("<http://ex/p1>"), None),
7526            (Some("<http://ex/s3>"), None, Some("<http://ex/o5>")),
7527            (None, Some("<http://ex/p1>"), Some("<http://ex/o5>")),
7528            (
7529                Some("<http://ex/s3>"),
7530                Some("<http://ex/p1>"),
7531                Some("<http://ex/o5>"),
7532            ),
7533        ];
7534
7535        for (image, tag) in [(&six, "six"), (&three, "three")] {
7536            let resident = Rete::open(image).unwrap();
7537            let leaked: &'static [u8] = Box::leak(image.clone().into_boxed_slice());
7538            let ranged = Rete::open_ranged(&SliceReader::new(leaked)).unwrap();
7539            let lazy =
7540                Rete::open_ranged_lazy_with(SliceReader::new(leaked), HEADER_LEN as u64).unwrap();
7541            for (s, p, o) in &shapes {
7542                let want: Vec<_> = {
7543                    let mut v = Rete::open(&six).unwrap().query(*s, *p, *o);
7544                    v.sort();
7545                    v
7546                };
7547                for (got, path) in [
7548                    (resident.query(*s, *p, *o), "resident"),
7549                    (ranged.query(*s, *p, *o), "ranged"),
7550                    (lazy.query(*s, *p, *o), "lazy"),
7551                ] {
7552                    let mut got = got;
7553                    got.sort();
7554                    assert_eq!(got, want, "{tag}/{path} disagreed on {:?}", (s, p, o));
7555                }
7556                // The named graph's container carries the same mask.
7557                let mut g = resident.query_in_graph(Some("<http://ex/g>"), *s, *p, *o);
7558                g.sort();
7559                assert_eq!(g, want, "{tag}/named disagreed on {:?}", (s, p, o));
7560            }
7561        }
7562
7563        // And the single-pattern ROUTED read (the `query-url` path) locates its
7564        // section by container position, not by the format's fixed six-wide
7565        // slot — the one place the two indexes differ.
7566        let leaked: &'static [u8] = Box::leak(three.clone().into_boxed_slice());
7567        let reader = SliceReader::new(leaked);
7568        let routed = Rete::query_ranged(&reader, None, None, Some("<http://ex/o5>")).unwrap();
7569        let mut want = Rete::open(&six)
7570            .unwrap()
7571            .query(None, None, Some("<http://ex/o5>"));
7572        want.sort();
7573        let mut got = routed;
7574        got.sort();
7575        assert_eq!(got, want, "routed single-pattern read on a lean file");
7576    }
7577
7578    #[test]
7579    fn read_coalesced_merges_within_gap_and_splits_beyond() {
7580        use crate::reader::{CountingReader, SliceReader};
7581        let bytes = vec![0u8; 4096];
7582        // Three 16-byte ranges: A..B gap = 32, B..C gap = 1024.
7583        let ranges = [
7584            ByteRange { offset: 0, len: 16 },
7585            ByteRange {
7586                offset: 48,
7587                len: 16,
7588            },
7589            ByteRange {
7590                offset: 1088,
7591                len: 16,
7592            },
7593        ];
7594        // Tight gap (16): nothing merges → one read per range.
7595        let r = CountingReader::new(SliceReader::new(&bytes));
7596        let out = read_coalesced(&r, &ranges, 16).unwrap();
7597        assert_eq!(out.len(), 3);
7598        assert_eq!(r.requests(), 3);
7599        // Gap 64 merges A+B (gap 32) but not C (gap 1024) → two reads.
7600        let r = CountingReader::new(SliceReader::new(&bytes));
7601        read_coalesced(&r, &ranges, 64).unwrap();
7602        assert_eq!(r.requests(), 2);
7603        // Gap 4096 merges all three into one read, over-fetching the gaps.
7604        let r = CountingReader::new(SliceReader::new(&bytes));
7605        read_coalesced(&r, &ranges, 4096).unwrap();
7606        assert_eq!(r.requests(), 1);
7607    }
7608
7609    fn build_image() -> Vec<u8> {
7610        let triples = [
7611            ("Alice", "knows", "Bob"),
7612            ("Bob", "knows", "Carol"),
7613            ("Alice", "age", "30"),
7614        ];
7615        let mut db = DictionaryBuilder::new();
7616        for (s, p, o) in triples {
7617            db.observe(s, p, o);
7618        }
7619        let dict = db.build();
7620
7621        let mut ib = GraphIndexBuilder::new();
7622        for (s, p, o) in triples {
7623            ib.push(dict.encode(s, p, o).unwrap());
7624        }
7625        let index = ib.build();
7626
7627        let (meta, levels) = build_pyramid_meta(&dict, &triples_ids(&dict), DEFAULT_TILE_BUDGET);
7628        write_file(&dict, &index, false, &meta, levels)
7629    }
7630
7631    fn triples_ids(dict: &Dictionary) -> Vec<(u32, u32, u32)> {
7632        [
7633            ("Alice", "knows", "Bob"),
7634            ("Bob", "knows", "Carol"),
7635            ("Alice", "age", "30"),
7636        ]
7637        .iter()
7638        .map(|(s, p, o)| dict.encode(s, p, o).unwrap())
7639        .collect()
7640    }
7641
7642    #[test]
7643    fn file_round_trips_header_and_counts() {
7644        let bytes = build_image();
7645        let rete = Rete::open(&bytes).unwrap();
7646        assert_eq!(rete.header().quad_count, 3);
7647        assert!(rete.header().term_count >= 5);
7648        let expected_codec = writer_codec();
7649        assert_eq!(rete.header().dict_codec, expected_codec);
7650        assert_eq!(rete.header().block_codec, expected_codec);
7651        assert_eq!(&bytes[bytes.len() - 4..], &MAGIC); // footer marker
7652    }
7653
7654    /// A file whose index was built with a tiny tile budget (forcing many
7655    /// tiles per permutation) must round-trip through write/open and answer
7656    /// every query shape identically — through both the in-memory and the
7657    /// routed ranged read paths.
7658    #[test]
7659    fn multi_tile_file_round_trips_and_routes() {
7660        let triples: Vec<(String, String, String)> = (0..200)
7661            .map(|i| {
7662                (
7663                    format!("<http://ex/s/{i}>"),
7664                    format!("<http://ex/p/{}>", i % 5),
7665                    format!("<http://ex/o/{}>", i % 23),
7666                )
7667            })
7668            .collect();
7669        let mut db = DictionaryBuilder::new();
7670        for (s, p, o) in &triples {
7671            db.observe(s, p, o);
7672        }
7673        let dict = db.build();
7674        let mut ib = GraphIndexBuilder::new().with_tile_budget(64);
7675        for (s, p, o) in &triples {
7676            ib.push(dict.encode(s, p, o).unwrap());
7677        }
7678        let index = ib.build();
7679        assert!(
7680            index.tile_sections()[0].len() > 3,
7681            "tiny budget must force many tiles"
7682        );
7683        let bytes = write_file(&dict, &index, false, &[], 0);
7684
7685        let rete = Rete::open(&bytes).unwrap();
7686        assert_eq!(rete.header().version, crate::header::LEGACY_FORMAT_VERSION);
7687        assert_eq!(rete.query(None, None, None).len(), 200);
7688        assert_eq!(rete.query(Some("<http://ex/s/7>"), None, None).len(), 1);
7689        assert_eq!(
7690            rete.query(None, Some("<http://ex/p/3>"), None).len(),
7691            40,
7692            "predicate extent spans tiles"
7693        );
7694        assert_eq!(
7695            rete.query(None, None, Some("<http://ex/o/22>")).len(),
7696            8 // 22, 45, 68, ... < 200
7697        );
7698
7699        // Routed ranged read must agree (and only decompress matching tiles).
7700        use crate::reader::SliceReader;
7701        let reader = SliceReader::new(&bytes);
7702        let routed = Rete::query_ranged(&reader, Some("<http://ex/s/7>"), None, None).unwrap();
7703        assert_eq!(routed.len(), 1);
7704        let routed = Rete::query_ranged(&reader, None, Some("<http://ex/p/3>"), None).unwrap();
7705        assert_eq!(routed.len(), 40);
7706        let routed = Rete::query_ranged(&reader, None, None, Some("<http://ex/o/22>")).unwrap();
7707        assert_eq!(routed.len(), 8);
7708    }
7709
7710    /// The tile-synopsis trailer round-trips through encode/parse, and each parsed
7711    /// synopsis is **exactly** the tile block's own b/c zone — so the directory
7712    /// can never prune a tile the tile itself would have matched.
7713    /// Section-internal byte offsets are u64: a directory whose tiles sit past
7714    /// 4 GiB must parse with exact offsets on EVERY platform. On wasm32 (32-bit
7715    /// usize) the old parse truncated a >4 GiB section length and rejected the
7716    /// tail ("dict chunk overruns section" on the first >4 GiB dictionary —
7717    /// crossref's 5.2 GB g.obj — the playground regression this guards).
7718    #[test]
7719    fn tile_directory_offsets_survive_past_4gib() {
7720        let mut dir = Vec::new();
7721        write_uvarint(&mut dir, 2); // two tiles
7722        write_uvarint(&mut dir, 5); // tile 1: Δmin_a
7723        write_uvarint(&mut dir, 0); //         span
7724        write_uvarint(&mut dir, 3 << 30); //   len = 3 GiB
7725        write_uvarint(&mut dir, 1); // tile 2: Δmin_a
7726        write_uvarint(&mut dir, 0);
7727        write_uvarint(&mut dir, 2 << 30); //   len = 2 GiB
7728        let total = dir.len() as u64 + (3u64 << 30) + (2u64 << 30) + 64;
7729        let entries = parse_tile_directory(&dir, total).unwrap();
7730        assert_eq!(entries.len(), 2);
7731        assert_eq!(entries[1].start, dir.len() as u64 + (3u64 << 30));
7732        assert!(
7733            entries[1].end > u32::MAX as u64,
7734            "tail tile sits past 4 GiB"
7735        );
7736        // a total smaller than the tiles must still reject the directory
7737        assert!(parse_tile_directory(&dir, 1 << 20).is_err());
7738    }
7739
7740    #[test]
7741    fn tile_synopsis_trailer_round_trips() {
7742        let mut ib = GraphIndexBuilder::new().with_tile_budget(64);
7743        for i in 0..200u32 {
7744            ib.push((i, i % 7, i % 13));
7745        }
7746        let index = ib.build();
7747        let tile_count = index.tile_sections()[0].len();
7748        assert!(tile_count > 3, "tiny budget forces many tiles");
7749
7750        let payload = encode_tiled_section(&index, 0, CODEC_NONE);
7751        let dir = parse_tile_directory(&payload, payload.len() as u64).unwrap();
7752        assert_eq!(dir.len(), tile_count);
7753        // The trailer sits past the last tile; the old directory parse stops there.
7754        let trailer_start = dir.iter().map(|e| e.end).max().unwrap();
7755        assert!(
7756            trailer_start < payload.len() as u64,
7757            "a trailer follows the tiles"
7758        );
7759        for e in &dir {
7760            assert!(
7761                e.end <= payload.len() as u64,
7762                "tiles still located within the payload"
7763            );
7764        }
7765        let syn = parse_tile_synopsis(&payload, trailer_start as usize, dir.len()).unwrap();
7766        for (e, (min_b, max_b, min_c, max_c)) in dir.iter().zip(syn) {
7767            let block = decompress(CODEC_NONE, &payload[e.start as usize..e.end as usize]).unwrap();
7768            let z = *crate::triples::TripleBlock::parse(&block).unwrap().zone();
7769            assert_eq!(
7770                (min_b, max_b, min_c, max_c),
7771                (z.min_b, z.max_b, z.min_c, z.max_c),
7772                "synopsis equals the tile's own zone"
7773            );
7774        }
7775    }
7776
7777    /// End-to-end safety: a synopsis-carrying file, opened **lazily** (range
7778    /// reads), must return exactly the brute-force answer for every pattern shape
7779    /// — the synopsis prune may never drop a real match.
7780    #[test]
7781    fn tile_synopsis_lazy_matches_reference_every_shape() {
7782        use crate::reader::{CountingReader, SliceReader};
7783        let triples: Vec<(String, String, String)> = (0..200u32)
7784            .map(|i| {
7785                (
7786                    format!("<http://ex/s/{i:04}>"),
7787                    format!("<http://ex/p/{}>", i % 7),
7788                    format!("<http://ex/o/{:04}>", i % 13),
7789                )
7790            })
7791            .collect();
7792        let mut db = DictionaryBuilder::new();
7793        for (s, p, o) in &triples {
7794            db.observe(s, p, o);
7795        }
7796        let dict = db.build();
7797        let mut ib = GraphIndexBuilder::new().with_tile_budget(64);
7798        for (s, p, o) in &triples {
7799            ib.push(dict.encode(s, p, o).unwrap());
7800        }
7801        let bytes = write_file(&dict, &ib.build(), false, &[], 0);
7802
7803        let eager = Rete::open(&bytes).unwrap();
7804        assert!(
7805            eager.header().has_tile_synopsis(),
7806            "new files set the synopsis flag"
7807        );
7808
7809        // `open_ranged_lazy` needs a `'static` reader; leak the image (the test
7810        // process exits straight after).
7811        let leaked: &'static [u8] = Box::leak(bytes.clone().into_boxed_slice());
7812        let reader = std::sync::Arc::new(CountingReader::new(SliceReader::new(leaked)));
7813        let lazy = Rete::open_ranged_lazy_with(reader, HEADER_LEN as u64).unwrap();
7814
7815        let brute = |s: Option<&str>, p: Option<&str>, o: Option<&str>| {
7816            let mut v: Vec<(String, String, String)> = triples
7817                .iter()
7818                .filter(|(a, b, c)| {
7819                    s.is_none_or(|x| x == a) && p.is_none_or(|x| x == b) && o.is_none_or(|x| x == c)
7820                })
7821                .cloned()
7822                .collect();
7823            v.sort();
7824            v
7825        };
7826        // Existing + absent terms in every position (and unbound) — 4×4×4 shapes.
7827        let sv = [
7828            None,
7829            Some("<http://ex/s/0007>"),
7830            Some("<http://ex/s/0130>"),
7831            Some("<http://ex/s/9999>"),
7832        ];
7833        let pv = [
7834            None,
7835            Some("<http://ex/p/3>"),
7836            Some("<http://ex/p/6>"),
7837            Some("<http://ex/p/999>"),
7838        ];
7839        let ov = [
7840            None,
7841            Some("<http://ex/o/0000>"),
7842            Some("<http://ex/o/0012>"),
7843            Some("<http://ex/o/9999>"),
7844        ];
7845        for &s in &sv {
7846            for &p in &pv {
7847                for &o in &ov {
7848                    let mut e = eager.query(s, p, o);
7849                    e.sort();
7850                    let mut l = lazy.query(s, p, o);
7851                    l.sort();
7852                    let r = brute(s, p, o);
7853                    assert_eq!(e, r, "eager {s:?} {p:?} {o:?}");
7854                    assert_eq!(l, r, "lazy {s:?} {p:?} {o:?} — synopsis over-pruned");
7855                }
7856            }
7857        }
7858        assert!(!lazy.index_incomplete(), "no lazy fetch failed");
7859    }
7860
7861    /// Byte-0 / polyglot experiment: a real `.rete` embedded behind a large HTML
7862    /// shell (so byte 0 is `<`, not `RETE`) still opens and queries LAZILY through
7863    /// an `OffsetReader`, touching only the graph's bytes and never the shell.
7864    #[test]
7865    fn polyglot_offset_reads_lazily() {
7866        use crate::reader::{
7867            detect_polyglot_base, CountingReader, OffsetReader, SliceReader, POLYGLOT_DIGITS,
7868            POLYGLOT_MARKER,
7869        };
7870
7871        // A real .rete image, hidden behind a 50 KB HTML shell that carries the
7872        // polyglot base-offset marker in its first bytes (a browser-ignored
7873        // comment). The .rete begins right after the shell.
7874        let image = build_image();
7875        let mut shell = Vec::new();
7876        shell.extend_from_slice(b"<!DOCTYPE html><html><head><!--");
7877        shell.extend_from_slice(POLYGLOT_MARKER);
7878        let digits_at = shell.len();
7879        shell.extend_from_slice(&[b'0'; POLYGLOT_DIGITS]); // patched below
7880        shell.extend_from_slice(b"--></head><body>a web page</body></html>\n");
7881        shell.resize(50_000, b' '); // >> the .rete, to prove the shell is untouched
7882        let base = shell.len() as u64;
7883        let digits = format!("{base:0width$}", width = POLYGLOT_DIGITS);
7884        shell[digits_at..digits_at + POLYGLOT_DIGITS].copy_from_slice(digits.as_bytes());
7885
7886        let mut poly = shell;
7887        poly.extend_from_slice(&image);
7888
7889        // byte 0 is a web page, not a .rete...
7890        assert_ne!(
7891            &poly[0..4],
7892            b"RETE",
7893            "polyglot must not start with the magic"
7894        );
7895        // ...but the marker in the first header window points at the embedded .rete.
7896        let detected = detect_polyglot_base(&poly[..crate::header::HEADER_LEN]).unwrap();
7897        assert_eq!(detected, base);
7898
7899        // Open it LAZILY through the offset shim and query it.
7900        let leaked: &'static [u8] = Box::leak(poly.into_boxed_slice());
7901        let counting = std::sync::Arc::new(CountingReader::new(SliceReader::new(leaked)));
7902        let lazy = Rete::open_ranged_lazy_with(
7903            OffsetReader::new(counting.clone(), detected),
7904            HEADER_LEN as u64,
7905        )
7906        .unwrap();
7907
7908        // The polyglot must answer exactly like the plain .rete it embeds.
7909        let plain = Rete::open(&image).unwrap();
7910        let mut want = plain.query(None, None, None);
7911        want.sort();
7912        let mut got = lazy.query(None, None, None);
7913        got.sort();
7914        assert_eq!(
7915            got, want,
7916            "polyglot lazy query differs from the plain .rete"
7917        );
7918        assert!(!got.is_empty(), "expected some triples");
7919        assert!(!lazy.index_incomplete(), "a lazy fetch failed");
7920
7921        // It was LAZY: it read fewer bytes than the HTML shell alone, i.e. it
7922        // never touched the prefix — only the embedded .rete's own bytes.
7923        let read = counting.bytes_read();
7924        eprintln!(
7925            "polyglot lazy read: HTML shell {base} B + .rete image {} B; \
7926             the query touched only {read} B (never the shell).",
7927            image.len()
7928        );
7929        assert!(
7930            read < base,
7931            "read {read} bytes; the HTML shell alone is {base} — the reader \
7932             touched the prefix instead of only the embedded .rete"
7933        );
7934    }
7935
7936    /// Build a small file whose objects are string literals, **with** a text
7937    /// index, and return `(image, triples)`. Shared by the text-index tests.
7938    #[cfg(test)]
7939    fn build_text_indexed(triples: &[(String, String, String)]) -> Vec<u8> {
7940        let mut db = DictionaryBuilder::new();
7941        for (s, p, o) in triples {
7942            db.observe(s, p, o);
7943        }
7944        let dict = db.build();
7945        let mut ib = GraphIndexBuilder::new().with_tile_budget(64);
7946        let mut id_triples: Vec<(u32, u32, u32)> = Vec::with_capacity(triples.len());
7947        for (s, p, o) in triples {
7948            let t = dict.encode(s, p, o).unwrap();
7949            ib.push(t);
7950            id_triples.push(t);
7951        }
7952        let index = ib.build();
7953        let text_index = compute_text_index(&dict, &id_triples);
7954        assert!(
7955            !text_index.is_empty(),
7956            "literals should produce a text index"
7957        );
7958        write_dataset_with_metadata(&dict, &index, &[], false, &[], 0, &[], &text_index)
7959    }
7960
7961    /// A `--text-index` build round-trips: `text_search` returns exactly the
7962    /// subjects whose literals contain the queried word(s), with AND across words
7963    /// and token-prefix — matching a brute-force scan of the literals.
7964    #[test]
7965    fn text_index_eager_matches_brute_force() {
7966        let triples: Vec<(String, String, String)> = vec![
7967            (
7968                "<http://ex/s0>",
7969                "<http://ex/label>",
7970                "\"alpha glucose phosphate\"",
7971            ),
7972            ("<http://ex/s1>", "<http://ex/label>", "\"beta Glucose\""),
7973            ("<http://ex/s2>", "<http://ex/label>", "\"gamma fructose\""),
7974            (
7975                "<http://ex/s3>",
7976                "<http://ex/note>",
7977                "\"einstein relativity\"",
7978            ),
7979            (
7980                "<http://ex/s4>",
7981                "<http://ex/ref>",
7982                "<http://ex/not-a-literal>",
7983            ),
7984        ]
7985        .into_iter()
7986        .map(|(s, p, o)| (s.to_string(), p.to_string(), o.to_string()))
7987        .collect();
7988        let bytes = build_text_indexed(&triples);
7989        let rete = Rete::open(&bytes).unwrap();
7990        assert!(rete.has_text_index());
7991
7992        // Brute-force reference: subjects whose literal objects contain all words.
7993        let brute = |words: &[&str]| -> Vec<String> {
7994            let mut v: Vec<String> = triples
7995                .iter()
7996                .filter(|(_, _, o)| {
7997                    crate::terms::is_literal(o)
7998                        && words.iter().all(|w| {
7999                            let wl = w.to_lowercase();
8000                            crate::terms::literal_lexical(o)
8001                                .unwrap()
8002                                .split(|c: char| !c.is_alphanumeric())
8003                                .any(|t| t.to_lowercase() == wl)
8004                        })
8005                })
8006                .map(|(s, _, _)| s.clone())
8007                .collect();
8008            v.sort();
8009            v.dedup();
8010            v
8011        };
8012
8013        let mut got = rete.text_search(&["glucose"], None, 0);
8014        got.sort();
8015        assert_eq!(got, brute(&["glucose"]), "case-insensitive single word");
8016
8017        // AND across two words: only s0 has both.
8018        let mut got = rete.text_search(&["glucose", "phosphate"], None, 0);
8019        got.sort();
8020        assert_eq!(got, brute(&["glucose", "phosphate"]));
8021
8022        // A word nobody has → empty.
8023        assert!(rete.text_search(&["zzznope"], None, 0).is_empty());
8024
8025        // Token-prefix: "ein…" matches "einstein".
8026        let got = rete.text_search(&[], Some("ein"), 0);
8027        assert_eq!(got, vec!["<http://ex/s3>".to_string()]);
8028
8029        // No text index → empty, has_text_index() false.
8030        let mut db = DictionaryBuilder::new();
8031        for (s, p, o) in &triples {
8032            db.observe(s, p, o);
8033        }
8034        let dict = db.build();
8035        let mut ib = GraphIndexBuilder::new();
8036        for (s, p, o) in &triples {
8037            ib.push(dict.encode(s, p, o).unwrap());
8038        }
8039        let plain = write_dataset(&dict, &ib.build(), &[], false, &[], 0);
8040        let plain_rete = Rete::open(&plain).unwrap();
8041        assert!(!plain_rete.has_text_index());
8042        assert!(plain_rete.text_search(&["glucose"], None, 0).is_empty());
8043    }
8044
8045    /// The lazy/remote path returns the same subjects as the eager path **and**
8046    /// faults only the token table + the queried posting list — never the whole
8047    /// postings blob. A `CountingReader` proves the byte budget stays small.
8048    #[test]
8049    fn text_index_lazy_faults_only_queried_postings() {
8050        use crate::reader::{CountingReader, SliceReader};
8051        // Many subjects so the postings blob is large relative to one posting:
8052        // every subject carries "common", but only a few carry "rare".
8053        let mut triples: Vec<(String, String, String)> = (0..300u32)
8054            .map(|i| {
8055                (
8056                    format!("<http://ex/s/{i:04}>"),
8057                    "<http://ex/label>".to_string(),
8058                    format!("\"common word number {i}\""),
8059                )
8060            })
8061            .collect();
8062        for i in [3u32, 77, 250] {
8063            triples.push((
8064                format!("<http://ex/s/{i:04}>"),
8065                "<http://ex/tag>".to_string(),
8066                "\"raretoken\"".to_string(),
8067            ));
8068        }
8069        let bytes = build_text_indexed(&triples);
8070        let eager = Rete::open(&bytes).unwrap();
8071        let mut want = eager.text_search(&["raretoken"], None, 0);
8072        want.sort();
8073        assert_eq!(want.len(), 3);
8074
8075        let leaked: &'static [u8] = Box::leak(bytes.clone().into_boxed_slice());
8076        let reader = std::sync::Arc::new(CountingReader::new(SliceReader::new(leaked)));
8077        let lazy = Rete::open_ranged_lazy_with(reader.clone(), HEADER_LEN as u64).unwrap();
8078        // Bytes pulled by the open itself (dict dirs, index dirs, named graphs) —
8079        // the text index is deferred and not touched yet.
8080        let before = reader.bytes_read();
8081        let mut got = lazy.text_search(&["raretoken"], None, 0);
8082        got.sort();
8083        assert_eq!(got, want, "lazy search matches eager");
8084        let pulled = reader.bytes_read() - before;
8085        // The search faulted the token table + the one "raretoken" posting; it must
8086        // be far less than the whole text-index section (300 "common" postings).
8087        let ti_len = eager.header().text_index_len;
8088        assert!(
8089            pulled < ti_len,
8090            "search pulled {pulled} B but the section is {ti_len} B — faulted too much"
8091        );
8092        assert!(!lazy.index_incomplete());
8093    }
8094
8095    /// `SearchView` answers exactly what a full open answers, having read
8096    /// strictly less: it skips the permutation tile directories and — the part
8097    /// that actually costs on a literal-heavy graph — the object-only
8098    /// dictionary's chunk directory.
8099    #[test]
8100    fn search_view_matches_full_open_for_far_fewer_bytes() {
8101        use crate::reader::{CountingReader, SliceReader};
8102        // Long object literals so the object-only dictionary (and its chunk
8103        // directory) dominates the file, as on a graph that stores abstracts.
8104        let filler = "lorem ipsum dolor sit amet consectetur adipiscing elit ".repeat(20);
8105        let mut triples: Vec<(String, String, String)> = (0..300u32)
8106            .map(|i| {
8107                (
8108                    format!("<http://ex/s/{i:04}>"),
8109                    "<http://ex/abstract>".to_string(),
8110                    format!("\"{filler} number {i}\""),
8111                )
8112            })
8113            .collect();
8114        for i in [3u32, 77, 250] {
8115            triples.push((
8116                format!("<http://ex/s/{i:04}>"),
8117                "<http://ex/tag>".to_string(),
8118                "\"raretoken\"".to_string(),
8119            ));
8120        }
8121        let bytes = build_text_indexed(&triples);
8122        let leaked: &'static [u8] = Box::leak(bytes.clone().into_boxed_slice());
8123
8124        let full_reader = std::sync::Arc::new(CountingReader::new(SliceReader::new(leaked)));
8125        let full = Rete::open_ranged_lazy_with(full_reader.clone(), HEADER_LEN as u64).unwrap();
8126        let mut want = full.text_search(&["raretoken"], None, 0);
8127        want.sort();
8128        assert_eq!(want.len(), 3);
8129
8130        let search_reader = std::sync::Arc::new(CountingReader::new(SliceReader::new(leaked)));
8131        let view = SearchView::open_ranged_with(search_reader.clone(), HEADER_LEN as u64).unwrap();
8132        assert!(view.has_text_index());
8133        let mut got = view.text_search(&["raretoken"], None, 0);
8134        got.sort();
8135        assert_eq!(got, want, "search view matches the full open");
8136        assert!(
8137            search_reader.bytes_read() < full_reader.bytes_read(),
8138            "search read {} B, full open read {} B — the narrow open bought nothing",
8139            search_reader.bytes_read(),
8140            full_reader.bytes_read()
8141        );
8142
8143        // A file without a text index says so from the header alone, and the
8144        // label-prefix mode still works through the same view.
8145        let mut prefix_hits = view.prefix_search("", 5);
8146        prefix_hits.sort();
8147        assert!(!prefix_hits.is_empty() || !view.has_pyramid());
8148    }
8149
8150    /// Nothing records a dictionary chunk directory's byte length, so a ranged
8151    /// open probes for it. That probe must cost about ONE directory, not several
8152    /// — it used to re-read the whole prefix on every round, so the sum was
8153    /// ~2x the last (already over-sized) read: 537 MB of range reads to fetch
8154    /// epfl-infoscience's 234 MB object-only directory. Here the same probe runs
8155    /// against a file whose object literals are long enough that its directory
8156    /// is well past the 8 KiB header prefix, and must (a) stay inside the
8157    /// append-only bound of 2x the directory and (b) beat what the old
8158    /// re-reading loop would have spent.
8159    #[test]
8160    fn dict_chunk_directory_probe_costs_about_one_directory() {
8161        use crate::reader::{CountingReader, SliceReader};
8162        // Long literals that share a 3 KiB prefix with their neighbour, each
8163        // ending in a 1 KiB tail of its own: the WKT-polygon / JSON-blob shape
8164        // (#198's weak case). The directory stores the SHORTEST SEPARATOR per
8165        // chunk, and a separator has to reproduce the whole shared prefix — so
8166        // these keys are still ~3 KiB and the directory still runs well past
8167        // the 8 KiB header prefix, which is what gives the probe something to
8168        // measure.
8169        //
8170        // This fixture used to put the varying bytes FIRST (`"{i:04} xxxx…"`).
8171        // That shape is the one separators annihilate: its keys collapse to ~4
8172        // bytes, the whole directory to 192 B, it fits the header prefix, and
8173        // the probe never runs — the assertion below then fails naming itself.
8174        // Keep the shared prefix.
8175        let head = "x".repeat(3000);
8176        let triples: Vec<(String, String, String)> = (0..2000u32)
8177            .map(|i| {
8178                let tail = char::from(b'a' + (i % 26) as u8).to_string().repeat(1000);
8179                (
8180                    format!("<http://ex/s/{i:04}>"),
8181                    "<http://ex/abstract>".to_string(),
8182                    format!("\"{head}{i:04}{tail}\""),
8183                )
8184            })
8185            .collect();
8186        let mut db = DictionaryBuilder::new();
8187        for (s, p, o) in &triples {
8188            db.observe(s, p, o);
8189        }
8190        let dict = db.build();
8191        let mut ib = GraphIndexBuilder::new();
8192        for (s, p, o) in &triples {
8193            ib.push(dict.encode(s, p, o).unwrap());
8194        }
8195        let bytes = write_file(&dict, &ib.build(), false, &[], 0);
8196
8197        let reader = CountingReader::new(SliceReader::new(&bytes));
8198        let header = Header::from_bytes(&reader.read_at(0, HEADER_LEN as u64).unwrap()).unwrap();
8199        // Section 2 is object-only: the one carrying the long literals.
8200        let section = locate_container_section_ranged(
8201            &reader,
8202            header.dictionary_offset,
8203            header.dictionary_len,
8204            2,
8205            4,
8206        )
8207        .unwrap();
8208
8209        let before = reader.bytes_read();
8210        let (_meta, entries) = read_dict_dir_ranged(&reader, section).unwrap();
8211        let spent = reader.bytes_read() - before;
8212
8213        // The first chunk body starts where the directory ends.
8214        let dir_end = entries[0].start;
8215        let head = reader.read_at(section.offset, 64).unwrap();
8216        let (header_len, n0) = read_uvarint(&head).unwrap();
8217        let dir_start = n0 as u64 + header_len;
8218        let dir_len = dir_end - dir_start;
8219        let dir_total = section.len - dir_start;
8220        assert!(
8221            dir_start + dir_len > 8192,
8222            "directory fits the header prefix ({dir_len} B) — the probe never runs"
8223        );
8224
8225        // What the previous loop spent: every round re-read from the start.
8226        let mut old = 0u64;
8227        let mut p = 4096u64.min(dir_total).max(1);
8228        loop {
8229            let got = p.min(dir_total);
8230            old += got;
8231            if got >= dir_len || p >= dir_total {
8232                break;
8233            }
8234            p = p.saturating_mul(2).min(dir_total);
8235        }
8236        assert!(
8237            spent <= dir_len * 2,
8238            "probe read {spent} B for a {dir_len} B directory — past the append-only bound"
8239        );
8240        assert!(
8241            spent < old,
8242            "probe read {spent} B; the re-reading loop it replaced spent {old} B"
8243        );
8244    }
8245
8246    // ---------------------------------------------------------------------
8247    // Separator routing keys (#198 option (d))
8248    // ---------------------------------------------------------------------
8249
8250    /// A verbatim-keyed twin of a chunked dictionary section: same header, same
8251    /// chunking, byte-identical compressed bodies — only every directory key
8252    /// replaced by the chunk's actual first term. That is exactly what
8253    /// `encode_chunked_dict_section` wrote before separator keys, so it is the
8254    /// control for every comparison below (and for the proptest).
8255    fn verbatim_keyed_twin(payload: &[u8], codec: u8) -> Vec<u8> {
8256        let (_meta, entries) = parse_chunked_dict_dir(payload, payload.len() as u64).unwrap();
8257        let (header_len, n0) = read_uvarint(payload).unwrap();
8258        let mut out = Vec::new();
8259        write_uvarint(&mut out, header_len);
8260        out.extend_from_slice(&payload[n0..n0 + header_len as usize]);
8261        write_uvarint(&mut out, entries.len() as u64);
8262        let mut prev_run = 0usize;
8263        for e in &entries {
8264            let body = decompress(codec, &payload[e.start as usize..e.end as usize]).unwrap();
8265            let first = crate::dict::run_first_term(&body, 0).unwrap();
8266            write_uvarint(&mut out, (e.first_run - prev_run) as u64);
8267            write_uvarint(&mut out, first.len() as u64);
8268            out.extend_from_slice(&first);
8269            write_uvarint(&mut out, e.end - e.start);
8270            prev_run = e.first_run;
8271        }
8272        for e in &entries {
8273            out.extend_from_slice(&payload[e.start as usize..e.end as usize]);
8274        }
8275        out
8276    }
8277
8278    /// Each chunk's (first term, last term), decoded from the bodies.
8279    fn chunk_bounds_of(payload: &[u8], codec: u8) -> Vec<(Vec<u8>, Vec<u8>)> {
8280        let (_meta, entries) = parse_chunked_dict_dir(payload, payload.len() as u64).unwrap();
8281        entries
8282            .iter()
8283            .map(|e| {
8284                let body = decompress(codec, &payload[e.start as usize..e.end as usize]).unwrap();
8285                (
8286                    crate::dict::run_first_term(&body, 0).unwrap(),
8287                    crate::dict::run_last_term(&body, 0, body.len()).unwrap(),
8288                )
8289            })
8290            .collect()
8291    }
8292
8293    /// A term set built to hit every shape the corpus has: plain IRIs, literals
8294    /// sharing a long prefix with their neighbour (separators are LONG there —
8295    /// the WKT/`ohm-full` weak case), literals that diverge in their first bytes
8296    /// (separators are tiny — the `epfl-infoscience` case), literals larger than
8297    /// `DICT_CHUNK_BUDGET` on their own, and multibyte UTF-8.
8298    fn adversarial_terms() -> Vec<String> {
8299        let mut terms: Vec<String> = Vec::new();
8300        for i in 0..6_000 {
8301            terms.push(format!("<http://example.org/work/{i:06}>"));
8302        }
8303        let shared = "X".repeat(900);
8304        for i in 0..4_000 {
8305            terms.push(format!("\"{shared}-{i:05}-tail\""));
8306        }
8307        let deep = "Y".repeat(3_000); // shared prefix far past any short key
8308        for i in 0..120 {
8309            terms.push(format!("\"{deep}{i:03}\""));
8310        }
8311        for i in 0..6 {
8312            // one literal larger than the 64 KiB budget: a chunk of one run
8313            terms.push(format!("\"BIG-{i:02}-{}\"", "Z".repeat(90_000)));
8314        }
8315        for i in 0..800 {
8316            terms.push(format!(
8317                "\"h\u{e9}llo-{}-\u{1F600}{i:04}\"",
8318                "\u{e9}".repeat(40)
8319            ));
8320        }
8321        for i in 0..2_000 {
8322            terms.push(format!("\"{i:05}-{}\"", "T".repeat(1_800)));
8323        }
8324        terms.sort();
8325        terms.dedup();
8326        terms
8327    }
8328
8329    fn chunked_section_of(terms: &[String]) -> Vec<u8> {
8330        let mut b = crate::dict::DictSectionBuilder::new().with_restart_interval(16);
8331        for t in terms {
8332            b.push(t.clone());
8333        }
8334        b.build()
8335    }
8336
8337    /// The invariant, asserted against the real chunk contents: the stored key
8338    /// of chunk `i` is strictly above chunk `i-1`'s LAST term and at most chunk
8339    /// `i`'s own first term, chunk 0's key is empty, and the keys ascend — the
8340    /// three facts `ChunkedSection::id`'s `partition_point` relies on.
8341    ///
8342    /// Plus the anti-regression the rename exists for: on a section like this
8343    /// the key must NOT be the first term. A writer that quietly went back to
8344    /// storing terms would pass every other test in this file — dumps and
8345    /// `term(id)` route by `first_run` and never read the key.
8346    #[test]
8347    fn chunk_directory_keys_are_separators_not_terms() {
8348        let terms = adversarial_terms();
8349        let raw = chunked_section_of(&terms);
8350        let payload = encode_chunked_dict_section(&raw, CODEC_NONE);
8351        let (_meta, entries) = parse_chunked_dict_dir(&payload, payload.len() as u64).unwrap();
8352        let bounds = chunk_bounds_of(&payload, CODEC_NONE);
8353        assert!(
8354            entries.len() > 20,
8355            "need a many-chunk section, got {}",
8356            entries.len()
8357        );
8358
8359        assert!(entries[0].key.is_empty(), "chunk 0 needs no separator");
8360        let mut key_bytes = 0usize;
8361        let mut term_bytes = 0usize;
8362        let mut differ = 0usize;
8363        for (i, e) in entries.iter().enumerate() {
8364            key_bytes += e.key.len();
8365            term_bytes += bounds[i].0.len();
8366            if e.key != bounds[i].0 {
8367                differ += 1;
8368            }
8369            assert!(
8370                e.key <= bounds[i].0,
8371                "chunk {i}: key must be <= its own first term"
8372            );
8373            if i > 0 {
8374                assert!(
8375                    e.key.as_slice() > bounds[i - 1].1.as_slice(),
8376                    "chunk {i}: key must be > the previous chunk's last term"
8377                );
8378                assert!(
8379                    e.key > entries[i - 1].key,
8380                    "chunk {i}: keys must ascend for partition_point"
8381                );
8382                // shortest: one byte past the divergence from the predecessor
8383                assert_eq!(
8384                    e.key,
8385                    crate::dict::shortest_separator(&bounds[i - 1].1, &bounds[i].0),
8386                    "chunk {i}: key is not the SHORTEST separator"
8387                );
8388            }
8389        }
8390        assert!(
8391            differ > 0,
8392            "every key equalled its first term — the writer stores terms again"
8393        );
8394        assert!(
8395            key_bytes * 4 < term_bytes,
8396            "separators bought almost nothing: {key_bytes} B vs {term_bytes} B of first terms"
8397        );
8398        eprintln!(
8399            "chunks={} first-term keys={term_bytes} B  separator keys={key_bytes} B ({:.1}x)",
8400            entries.len(),
8401            term_bytes as f64 / key_bytes.max(1) as f64
8402        );
8403    }
8404
8405    /// A separator-keyed section answers **identically** to the verbatim-keyed
8406    /// one built over the same bodies, under the same unmodified reader: every
8407    /// id, every term, and every boundary probe — the first and last term of
8408    /// each chunk, one byte either side of each, strings that fall strictly
8409    /// *between* two chunks, truncations, the empty string, and probes below
8410    /// and above the whole section.
8411    ///
8412    /// Probing `id(term)` is the entire point. A broken key (a truncation, a
8413    /// hash) leaves `term(id)`, `dump` and `export` byte-perfect — they route by
8414    /// `first_run` — and only lookups lie.
8415    #[test]
8416    fn separator_keys_route_identically_to_verbatim_first_terms() {
8417        let terms = adversarial_terms();
8418        let raw = chunked_section_of(&terms);
8419        let sep_payload = encode_chunked_dict_section(&raw, CODEC_NONE);
8420        let ver_payload = verbatim_keyed_twin(&sep_payload, CODEC_NONE);
8421
8422        // Same chunking, byte-identical bodies — only the keys differ.
8423        let (_ms, es) = parse_chunked_dict_dir(&sep_payload, sep_payload.len() as u64).unwrap();
8424        let (_mv, ev) = parse_chunked_dict_dir(&ver_payload, ver_payload.len() as u64).unwrap();
8425        assert_eq!(es.len(), ev.len());
8426        for (a, b) in es.iter().zip(ev.iter()) {
8427            assert_eq!(a.first_run, b.first_run);
8428            assert_eq!(
8429                &sep_payload[a.start as usize..a.end as usize],
8430                &ver_payload[b.start as usize..b.end as usize],
8431                "chunk bodies must be identical"
8432            );
8433        }
8434        assert!(
8435            sep_payload.len() < ver_payload.len(),
8436            "separator payload is not smaller"
8437        );
8438
8439        let sec_s = decode_chunked_dict_section(
8440            &sep_payload,
8441            CODEC_NONE,
8442            crate::chunk_cache::ChunkCache::unlimited_arc(),
8443            0,
8444        )
8445        .unwrap();
8446        let sec_v = decode_chunked_dict_section(
8447            &ver_payload,
8448            CODEC_NONE,
8449            crate::chunk_cache::ChunkCache::unlimited_arc(),
8450            0,
8451        )
8452        .unwrap();
8453
8454        // 1. every present term → the same, correct id
8455        for (i, t) in terms.iter().enumerate() {
8456            let want = Some(i as u32 + 1);
8457            assert_eq!(sec_v.id(t), want, "verbatim id() lost a term");
8458            assert_eq!(sec_s.id(t), want, "separator id() lost a term");
8459        }
8460        // 2. every id → the same term
8461        for i in 0..terms.len() as u32 {
8462            assert_eq!(sec_s.term(i + 1), sec_v.term(i + 1));
8463            assert_eq!(
8464                sec_s.term(i + 1).as_deref(),
8465                Some(terms[i as usize].as_str())
8466            );
8467        }
8468        // 3. boundary probes, from the real chunk bounds
8469        let bounds = chunk_bounds_of(&sep_payload, CODEC_NONE);
8470        let mut probes: Vec<String> = Vec::new();
8471        let push = |bytes: &[u8], probes: &mut Vec<String>| {
8472            if let Ok(s) = std::str::from_utf8(bytes) {
8473                probes.push(s.to_string());
8474            }
8475        };
8476        for (first, last) in &bounds {
8477            for base in [first, last] {
8478                push(base, &mut probes); // present
8479                let mut after = base.clone();
8480                after.push(0x01); // immediately after — between chunks when base = last
8481                push(&after, &mut probes);
8482                let mut far = base.clone();
8483                far.push(b'~');
8484                push(&far, &mut probes);
8485                if !base.is_empty() {
8486                    push(&base[..base.len() - 1], &mut probes); // truncated
8487                    let mut down = base.clone();
8488                    *down.last_mut().unwrap() = down.last().unwrap().wrapping_sub(1);
8489                    push(&down, &mut probes);
8490                    let mut up = base.clone();
8491                    *up.last_mut().unwrap() = up.last().unwrap().wrapping_add(1);
8492                    push(&up, &mut probes);
8493                }
8494            }
8495        }
8496        // 4. the stored separators themselves, and the ends of the world
8497        for e in &es {
8498            push(&e.key, &mut probes);
8499            let mut plus = e.key.clone();
8500            plus.push(0);
8501            push(&plus, &mut probes);
8502        }
8503        probes.push(String::new()); // the empty key, i.e. chunk 0's
8504        probes.push("\u{1}".to_string()); // below every term
8505        probes.push("\u{10FFFF}\u{10FFFF}".to_string()); // above every term
8506        probes.push(terms.first().unwrap().clone());
8507        probes.push(terms.last().unwrap().clone());
8508
8509        let mut hits = 0usize;
8510        for p in &probes {
8511            let (a, b) = (sec_v.id(p), sec_s.id(p));
8512            assert_eq!(
8513                a,
8514                b,
8515                "probe diverged: {:?}",
8516                p.get(..p.len().min(48)).unwrap_or(p.as_str())
8517            );
8518            hits += usize::from(a.is_some());
8519        }
8520        assert!(
8521            probes.len() > 500 && hits > 50,
8522            "probe set is too thin: {} probes, {hits} of them present",
8523            probes.len()
8524        );
8525        eprintln!(
8526            "{} boundary probes, {hits} present, all identical",
8527            probes.len()
8528        );
8529    }
8530
8531    /// The same claim as a **property**: for arbitrary term sets and arbitrary
8532    /// probes, a separator-keyed section and the verbatim-keyed twin built from
8533    /// the same input are indistinguishable through the reader. The fixed test
8534    /// above picks the shapes I thought of; this one shrinks a counterexample
8535    /// out of the ones I did not.
8536    mod separator_props {
8537        use super::*;
8538        use proptest::prelude::*;
8539
8540        /// Sorted, deduplicated term sets mixing the three shapes that decide a
8541        /// separator's length: short IRIs, literals whose payload comes BEFORE
8542        /// the discriminating bytes (long shared prefixes ⇒ long separators),
8543        /// and literals that diverge immediately (⇒ one-byte separators). Sizes
8544        /// straddle `DICT_CHUNK_BUDGET`, so single- and many-chunk sections both
8545        /// occur.
8546        fn term_set() -> impl Strategy<Value = Vec<String>> {
8547            prop::collection::vec(("[a-c]{1,6}", 0u8..3, 300usize..1200), 60..250).prop_map(
8548                |specs| {
8549                    let mut terms: Vec<String> = specs
8550                        .into_iter()
8551                        .enumerate()
8552                        .map(|(i, (core, kind, pad))| {
8553                            let block = "z".repeat(pad);
8554                            match kind {
8555                                0 => format!("<http://ex/{core}/{i:05}>"),
8556                                1 => format!("\"{block}{core}{i:05}\""),
8557                                _ => format!("\"{core}{i:05}{block}\""),
8558                            }
8559                        })
8560                        .collect();
8561                    terms.sort();
8562                    terms.dedup();
8563                    terms
8564                },
8565            )
8566        }
8567
8568        proptest! {
8569            #[test]
8570            fn prop_separator_keyed_section_answers_like_verbatim(
8571                terms in term_set(),
8572                extra in prop::collection::vec(any::<String>(), 0..8),
8573            ) {
8574                let raw = chunked_section_of(&terms);
8575                let sep = encode_chunked_dict_section(&raw, CODEC_NONE);
8576                let ver = verbatim_keyed_twin(&sep, CODEC_NONE);
8577                prop_assert!(sep.len() <= ver.len());
8578
8579                // The separator invariant, against the real chunk contents.
8580                let (_m, es) = parse_chunked_dict_dir(&sep, sep.len() as u64).unwrap();
8581                let bounds = chunk_bounds_of(&sep, CODEC_NONE);
8582                prop_assert!(es[0].key.is_empty());
8583                for i in 1..es.len() {
8584                    prop_assert!(es[i].key.as_slice() > bounds[i - 1].1.as_slice());
8585                    prop_assert!(es[i].key <= bounds[i].0);
8586                    prop_assert!(es[i].key > es[i - 1].key);
8587                }
8588
8589                let sec_s = decode_chunked_dict_section(
8590                    &sep,
8591                    CODEC_NONE,
8592                    crate::chunk_cache::ChunkCache::unlimited_arc(),
8593                    0,
8594                )
8595                .unwrap();
8596                let sec_v = decode_chunked_dict_section(
8597                    &ver,
8598                    CODEC_NONE,
8599                    crate::chunk_cache::ChunkCache::unlimited_arc(),
8600                    0,
8601                )
8602                .unwrap();
8603
8604                for (i, t) in terms.iter().enumerate() {
8605                    prop_assert_eq!(sec_s.id(t), Some(i as u32 + 1));
8606                    prop_assert_eq!(sec_v.id(t), Some(i as u32 + 1));
8607                    prop_assert_eq!(sec_s.term(i as u32 + 1), sec_v.term(i as u32 + 1));
8608                }
8609
8610                // Probes: arbitrary strings, plus mutations of every boundary.
8611                let mut probes: Vec<String> = extra;
8612                probes.push(String::new());
8613                for (first, last) in &bounds {
8614                    for base in [first, last] {
8615                        if let Ok(s) = std::str::from_utf8(base) {
8616                            probes.push(s.to_string());
8617                            probes.push(format!("{s}\u{1}"));
8618                            probes.push(s[..s.len() - s.chars().next_back()
8619                                .map(char::len_utf8).unwrap_or(0)].to_string());
8620                        }
8621                    }
8622                }
8623                for e in &es {
8624                    if let Ok(s) = std::str::from_utf8(&e.key) {
8625                        probes.push(s.to_string());
8626                    }
8627                }
8628                for p in &probes {
8629                    prop_assert_eq!(sec_s.id(p), sec_v.id(p), "probe diverged: {:?}", p);
8630                }
8631            }
8632        }
8633    }
8634
8635    /// The TEXT_INDEX section is inside the content hash: a freshly built
8636    /// text-indexed file must pass `verify()`, and flipping a byte inside the
8637    /// section must break it. (Regression: `verify()` once rebuilt the hash
8638    /// without the text index, so every `--text-index` file failed as corrupt.)
8639    #[test]
8640    fn text_index_is_tamper_evident_and_verifies() {
8641        let triples: Vec<(String, String, String)> = vec![(
8642            "<http://ex/s0>".to_string(),
8643            "<http://ex/label>".to_string(),
8644            "\"alpha glucose phosphate\"".to_string(),
8645        )];
8646        let bytes = build_text_indexed(&triples);
8647        let header = Rete::open(&bytes).unwrap().header().clone();
8648        assert!(header.text_index_len > 0);
8649        assert!(verify(&bytes).unwrap(), "a text-indexed build must verify");
8650
8651        let mut tampered = bytes.clone();
8652        tampered[header.text_index_offset as usize] ^= 0xff;
8653        assert!(
8654            !verify(&tampered).unwrap(),
8655            "tampering with the text index must break verify()"
8656        );
8657    }
8658
8659    /// End-to-end win: on a remote (range-read) file, a lookup whose routed tile
8660    /// is ruled out by a bound secondary fetches **fewer bytes** with the synopsis
8661    /// than without it — and the answer is identical (empty) either way.
8662    #[test]
8663    fn synopsis_cuts_remote_fetch_bytes() {
8664        use crate::header::FLAG_TILE_SYNOPSIS;
8665        use crate::reader::{CountingReader, SliceReader};
8666
8667        // Zero-padded terms ⇒ dictionary ids are monotonic in i; subject s_i pairs
8668        // only with object o_i, so an OSP tile (routed by object) holds a
8669        // contiguous subject range — a subject from a far tile is provably absent.
8670        let triples: Vec<(String, String, String)> = (0..400u32)
8671            .map(|i| {
8672                (
8673                    format!("<http://ex/s/{i:04}>"),
8674                    "<http://ex/p>".to_string(),
8675                    format!("<http://ex/o/{i:04}>"),
8676                )
8677            })
8678            .collect();
8679        let mut db = DictionaryBuilder::new();
8680        for (s, p, o) in &triples {
8681            db.observe(s, p, o);
8682        }
8683        let dict = db.build();
8684        let mut ib = GraphIndexBuilder::new().with_tile_budget(64);
8685        for (s, p, o) in &triples {
8686            ib.push(dict.encode(s, p, o).unwrap());
8687        }
8688        let bytes = write_file(&dict, &ib.build(), false, &[], 0);
8689
8690        // (s_0395, ?, o_0005): routes OSP by the (early) object, secondary = the
8691        // (late) subject — outside that tile's subject range, so the synopsis
8692        // prunes the one routed tile.
8693        let q = (Some("<http://ex/s/0395>"), None, Some("<http://ex/o/0005>"));
8694        // Measure the bytes the QUERY pulls (after open) — isolating the per-query
8695        // saving from the one-time synopsis trailer reads done at open, which a
8696        // persistent remote session amortizes over many queries.
8697        let query_bytes = |image: &[u8]| -> (u64, usize) {
8698            let leaked: &'static [u8] = Box::leak(image.to_vec().into_boxed_slice());
8699            let reader = std::sync::Arc::new(CountingReader::new(SliceReader::new(leaked)));
8700            let rete = Rete::open_ranged_lazy_with(reader.clone(), HEADER_LEN as u64).unwrap();
8701            let before = reader.bytes_read(); // after open (incl. trailer reads)
8702            let n = rete.query(q.0, q.1, q.2).len();
8703            assert!(!rete.index_incomplete());
8704            (reader.bytes_read() - before, n)
8705        };
8706
8707        let (on_bytes, on_n) = query_bytes(&bytes);
8708        // Same file with the synopsis flag cleared = an older reader's behavior.
8709        let mut off = bytes.clone();
8710        off[5] &= !FLAG_TILE_SYNOPSIS;
8711        let (off_bytes, off_n) = query_bytes(&off);
8712
8713        assert_eq!(on_n, 0, "the pair never co-occurs");
8714        assert_eq!(off_n, 0, "same answer without the synopsis");
8715        // Both pay the same dictionary-resolution bytes; the difference is the one
8716        // routed index tile that the synopsis skips (and the no-synopsis path
8717        // fetches only to have its zone map reject it).
8718        assert!(
8719            on_bytes < off_bytes,
8720            "synopsis skips the routed tile fetch: {on_bytes} < {off_bytes}"
8721        );
8722    }
8723
8724    /// A double-bound-object intersection (`?p P o1 ; P o2 ; label ?l`) — the
8725    /// shape whose REMOTE join strategy changed (scan + hash-join instead of
8726    /// probing each prefix row) — must return the SAME rows opened eagerly (in
8727    /// memory) and lazily (remote-style, `is_remote()` true). Strategy is a
8728    /// performance choice; the result multiset is invariant.
8729    #[test]
8730    fn double_bound_object_join_eager_matches_lazy() {
8731        use crate::reader::SliceReader;
8732        let occ = "<http://ex/occ>";
8733        let phys = "<http://ex/physicist>";
8734        let phil = "<http://ex/philosopher>";
8735        let label = "<http://www.w3.org/2000/01/rdf-schema#label>";
8736        // p00..p19 are physicists; p00..p09 are also philosophers (the answer).
8737        let mut triples: Vec<(String, String, String)> = Vec::new();
8738        for i in 0..20u32 {
8739            triples.push((format!("<http://ex/p/{i:02}>"), occ.into(), phys.into()));
8740            if i < 10 {
8741                triples.push((format!("<http://ex/p/{i:02}>"), occ.into(), phil.into()));
8742            }
8743            triples.push((
8744                format!("<http://ex/p/{i:02}>"),
8745                label.into(),
8746                format!("\"Name {i:02}\""),
8747            ));
8748        }
8749        let mut db = DictionaryBuilder::new();
8750        for (s, p, o) in &triples {
8751            db.observe(s, p, o);
8752        }
8753        let dict = db.build();
8754        let mut ib = GraphIndexBuilder::new().with_tile_budget(16);
8755        for (s, p, o) in &triples {
8756            ib.push(dict.encode(s, p, o).unwrap());
8757        }
8758        let bytes = write_file(&dict, &ib.build(), false, &[], 0);
8759
8760        let q = "SELECT ?l WHERE { \
8761            ?p <http://ex/occ> <http://ex/physicist> ; \
8762               <http://ex/occ> <http://ex/philosopher> ; \
8763               <http://www.w3.org/2000/01/rdf-schema#label> ?l }";
8764        let run = |rete: &Rete| -> Vec<String> {
8765            let (_, sols) = crate::eval_sparql(rete, q).unwrap();
8766            let mut v: Vec<String> = sols.iter().filter_map(|b| b.get("l").cloned()).collect();
8767            v.sort();
8768            v
8769        };
8770
8771        let eager_rows = run(&Rete::open(&bytes).unwrap());
8772        let leaked: &'static [u8] = Box::leak(bytes.clone().into_boxed_slice());
8773        let lazy = Rete::open_ranged_lazy_with(
8774            std::sync::Arc::new(SliceReader::new(leaked)),
8775            HEADER_LEN as u64,
8776        )
8777        .unwrap();
8778        let lazy_rows = run(&lazy);
8779        assert!(!lazy.index_incomplete());
8780
8781        assert_eq!(eager_rows.len(), 10, "the 10 physicist∩philosopher labels");
8782        assert_eq!(eager_rows, lazy_rows, "eager and lazy must agree exactly");
8783    }
8784
8785    /// A dictionary big enough to split into multiple chunks per section must
8786    /// round-trip every id↔term mapping through the chunked (v0.2) encoding —
8787    /// including terms at chunk boundaries and absent near-misses.
8788    #[test]
8789    fn multi_chunk_dictionary_round_trips() {
8790        let mut db = DictionaryBuilder::new();
8791        let term = |i: u32| format!("<http://example.org/some/long/prefix/entity/{i:06}>");
8792        for i in 0..6000u32 {
8793            db.observe(&term(i), "<http://ex/p>", &term(i + 1));
8794        }
8795        let dict = db.build();
8796        let mut ib = GraphIndexBuilder::new();
8797        for i in 0..6000u32 {
8798            ib.push(
8799                dict.encode(&term(i), "<http://ex/p>", &term(i + 1))
8800                    .unwrap(),
8801            );
8802        }
8803        let bytes = write_file(&dict, &ib.build(), false, &[], 0);
8804        let rete = Rete::open(&bytes).unwrap();
8805        let d = rete.dictionary();
8806        assert_eq!(d.term_count(), dict.term_count());
8807        for i in (0..6000).step_by(97).chain([0, 1, 5999, 6000]) {
8808            let t = term(i);
8809            let sid = dict.subject_id(&t);
8810            assert_eq!(d.subject_id(&t), sid, "subject_id({t})");
8811            if let Some(id) = sid {
8812                assert_eq!(d.subject_term(id).as_deref(), Some(t.as_str()));
8813            }
8814            let oid = dict.object_id(&t);
8815            assert_eq!(d.object_id(&t), oid, "object_id({t})");
8816        }
8817        assert_eq!(d.subject_id("<http://example.org/absent>"), None);
8818        assert_eq!(d.predicate_id("<http://ex/p>"), Some(1));
8819        assert_eq!(d.predicate_term(1).as_deref(), Some("<http://ex/p>"));
8820    }
8821
8822    #[test]
8823    #[cfg(feature = "compression")]
8824    fn compression_shrinks_repetitive_data() {
8825        // Many triples sharing IRI prefixes — exactly what front-coding + zstd
8826        // should crush. The compressed file must be far smaller than the raw
8827        // term bytes, and still query correctly.
8828        let mut db = DictionaryBuilder::new();
8829        let triples: Vec<(String, String, String)> = (0..500)
8830            .map(|i| {
8831                (
8832                    format!("<http://example.org/entity/{i}>"),
8833                    "<http://example.org/p/relatedTo>".to_string(),
8834                    format!("<http://example.org/entity/{}>", (i + 1) % 500),
8835                )
8836            })
8837            .collect();
8838        for (s, p, o) in &triples {
8839            db.observe(s, p, o);
8840        }
8841        let dict = db.build();
8842        let mut ib = GraphIndexBuilder::new();
8843        for (s, p, o) in &triples {
8844            ib.push(dict.encode(s, p, o).unwrap());
8845        }
8846        let bytes = write_file(&dict, &ib.build(), false, &[], 0);
8847
8848        let raw: usize = triples
8849            .iter()
8850            .map(|(s, p, o)| s.len() + p.len() + o.len())
8851            .sum();
8852        assert!(
8853            bytes.len() < raw / 2,
8854            "expected strong compression: file {} vs raw terms {raw}",
8855            bytes.len()
8856        );
8857
8858        // Still queryable after compression.
8859        let rete = Rete::open(&bytes).unwrap();
8860        let r = rete.query(Some("<http://example.org/entity/0>"), None, None);
8861        assert_eq!(r.len(), 1);
8862        assert_eq!(r[0].2, "<http://example.org/entity/1>");
8863    }
8864
8865    fn big_file_with_pyramid() -> Vec<u8> {
8866        // A ring of 300 entities -> index dwarfs dict+meta.
8867        let triples: Vec<(String, String, String)> = (0..300)
8868            .map(|i| {
8869                (
8870                    format!("<http://ex/e{i}>"),
8871                    "<http://ex/next>".to_string(),
8872                    format!("<http://ex/e{}>", (i + 1) % 300),
8873                )
8874            })
8875            .collect();
8876        let mut db = DictionaryBuilder::new();
8877        for (s, p, o) in &triples {
8878            db.observe(s, p, o);
8879        }
8880        let dict = db.build();
8881        let ids: Vec<_> = triples
8882            .iter()
8883            .map(|(s, p, o)| dict.encode(s, p, o).unwrap())
8884            .collect();
8885        let mut ib = GraphIndexBuilder::new();
8886        for &t in &ids {
8887            ib.push(t);
8888        }
8889        let (meta, levels) = build_pyramid_meta(&dict, &ids, DEFAULT_TILE_BUDGET);
8890        write_file(&dict, &ib.build(), false, &meta, levels)
8891    }
8892
8893    #[test]
8894    fn ranged_open_is_minimal_and_correct() {
8895        use crate::reader::{CountingReader, SliceReader};
8896        let bytes = big_file_with_pyramid();
8897
8898        let full = CountingReader::new(SliceReader::new(&bytes));
8899        let rete = Rete::open_ranged(&full).unwrap();
8900        // Full open touches at most 4 ranges (header, dict, index, meta).
8901        assert!(full.requests() <= 4, "requests = {}", full.requests());
8902        assert_eq!(
8903            rete.query(Some("<http://ex/e0>"), None, None)[0].2,
8904            "<http://ex/e1>"
8905        );
8906
8907        // Summary-only open skips the index → strictly fewer bytes than the file.
8908        let summ_reader = CountingReader::new(SliceReader::new(&bytes));
8909        let view = SummaryView::open_ranged(&summ_reader).unwrap().unwrap();
8910        assert!(!view.summary.is_empty());
8911        assert!(
8912            summ_reader.bytes_read() < bytes.len() as u64,
8913            "summary read {} of {} bytes",
8914            summ_reader.bytes_read(),
8915            bytes.len()
8916        );
8917        // And fewer than a full open, since it never fetched the index.
8918        assert!(summ_reader.bytes_read() < full.bytes_read());
8919    }
8920
8921    #[test]
8922    fn content_hash_is_set_and_verifies() {
8923        let bytes = build_image();
8924        let rete = Rete::open(&bytes).unwrap();
8925        assert_ne!(
8926            rete.header().content_hash,
8927            [0u8; 16],
8928            "hash must be populated"
8929        );
8930        assert!(verify(&bytes).unwrap(), "freshly built file verifies");
8931        assert!(
8932            verify_ranged(&crate::reader::SliceReader::new(&bytes)).unwrap(),
8933            "the streaming verify agrees"
8934        );
8935
8936        // Same data builds an identical hash (deterministic).
8937        assert_eq!(
8938            Rete::open(&build_image()).unwrap().header().content_hash,
8939            rete.header().content_hash
8940        );
8941
8942        // Corrupting a payload byte breaks verification.
8943        let mut tampered = bytes.clone();
8944        let last = tampered.len() - 5; // inside payload, before footer magic
8945        tampered[last] ^= 0xff;
8946        assert!(!verify(&tampered).unwrap());
8947        assert!(!verify_ranged(&crate::reader::SliceReader::new(&tampered)).unwrap());
8948        // A truncated file is an error, never a pass.
8949        let short = &bytes[..bytes.len() / 2];
8950        assert!(verify_ranged(&crate::reader::SliceReader::new(short)).is_err());
8951        assert!(verify_ranged(&crate::reader::SliceReader::new(&bytes[..10])).is_err());
8952    }
8953
8954    /// Build the standard 3-triple image with an opaque metadata payload.
8955    fn build_with_metadata(meta: &[u8]) -> Vec<u8> {
8956        let triples = [
8957            ("Alice", "knows", "Bob"),
8958            ("Bob", "knows", "Carol"),
8959            ("Alice", "age", "30"),
8960        ];
8961        let mut db = DictionaryBuilder::new();
8962        for (s, p, o) in triples {
8963            db.observe(s, p, o);
8964        }
8965        let dict = db.build();
8966        let ids: Vec<_> = triples
8967            .iter()
8968            .map(|(s, p, o)| dict.encode(s, p, o).unwrap())
8969            .collect();
8970        let mut ib = GraphIndexBuilder::new();
8971        for &t in &ids {
8972            ib.push(t);
8973        }
8974        let (pmeta, levels) = build_pyramid_meta(&dict, &ids, DEFAULT_TILE_BUDGET);
8975        write_dataset_with_metadata(&dict, &ib.build(), &[], false, &pmeta, levels, meta, &[])
8976    }
8977
8978    #[test]
8979    fn metadata_round_trips_and_shifts_offsets() {
8980        let card = br#"{"title":"My Dataset"}"#;
8981        let bytes = build_with_metadata(card);
8982        let rete = Rete::open(&bytes).unwrap();
8983
8984        // The opaque payload reads back verbatim.
8985        assert_eq!(rete.metadata(), Some(card.as_slice()));
8986        let h = rete.header();
8987        assert_eq!(h.metadata_offset, HEADER_LEN as u64);
8988        assert_eq!(h.metadata_len, card.len() as u64);
8989        // The dictionary (and everything after it) shifted forward by the card.
8990        assert_eq!(h.dictionary_offset, HEADER_LEN as u64 + card.len() as u64);
8991
8992        // The index still decodes correctly at its shifted offset.
8993        assert_eq!(
8994            rete.query(Some("Bob"), Some("knows"), Some("Carol")).len(),
8995            1
8996        );
8997        // The card is inside the content hash, so the file still verifies.
8998        assert!(verify(&bytes).unwrap());
8999    }
9000
9001    #[test]
9002    fn empty_metadata_is_byte_identical_to_plain_writer() {
9003        // The `&[]` path must produce exactly the bytes of the metadata-free
9004        // writer for identical inputs — old files and outputs are unchanged.
9005        assert_eq!(
9006            build_with_metadata(&[]),
9007            build_image(),
9008            "empty-metadata output must equal the plain writer byte-for-byte"
9009        );
9010    }
9011
9012    /// The contract `replace_metadata` has to meet to be usable at the end of a
9013    /// build: swapping the card must land on **exactly** the bytes the writer
9014    /// would have produced with that card in the first place — shorter, longer,
9015    /// and gone — so a spliced file and a rebuilt one are the same file.
9016    #[test]
9017    fn replace_metadata_matches_a_direct_write() {
9018        let first = br#"{"title":"My Dataset","queries":[1,2,3]}"#;
9019        let bytes = build_with_metadata(first);
9020        for target in [
9021            br#"{"title":"My Dataset","queries":[1,2]}"#.as_slice(), // shorter
9022            br#"{"title":"My Dataset","queries":[1,2,3,4,5,6,7]}"#.as_slice(), // longer
9023            b"".as_slice(),                                          // removed
9024            first.as_slice(),                                        // unchanged
9025        ] {
9026            let spliced = replace_metadata(&bytes, target).unwrap();
9027            assert_eq!(
9028                spliced,
9029                build_with_metadata(target),
9030                "splicing {} bytes of metadata must equal a direct write",
9031                target.len()
9032            );
9033            assert!(
9034                verify(&spliced).unwrap(),
9035                "the new hash covers the new card"
9036            );
9037            let rete = Rete::open(&spliced).unwrap();
9038            assert_eq!(
9039                rete.metadata(),
9040                (!target.is_empty()).then_some(target),
9041                "the new payload reads back verbatim"
9042            );
9043            // The data is untouched by the surgery.
9044            assert_eq!(
9045                rete.query(Some("Bob"), Some("knows"), Some("Carol")).len(),
9046                1
9047            );
9048        }
9049    }
9050
9051    /// A build-info section survives the splice, still adjacent to the card, and
9052    /// still outside the content hash.
9053    #[test]
9054    fn replace_metadata_keeps_build_info_adjacent() {
9055        let with_info =
9056            attach_build_info(&build_with_metadata(br#"{"a":1}"#), br#"{"builder":"x"}"#).unwrap();
9057        let spliced = replace_metadata(&with_info, br#"{"a":1,"b":2}"#).unwrap();
9058        let h = Header::from_bytes(&spliced).unwrap();
9059        assert_eq!(h.metadata_offset, HEADER_LEN as u64);
9060        assert_eq!(h.build_info_offset, HEADER_LEN as u64 + h.metadata_len);
9061        assert_eq!(
9062            read_build_info(&spliced).unwrap().as_deref(),
9063            Some(br#"{"builder":"x"}"#.as_slice())
9064        );
9065        assert!(verify(&spliced).unwrap());
9066        // Same card, same content hash, whatever the build-info says.
9067        assert_eq!(
9068            h.content_hash,
9069            Header::from_bytes(&build_with_metadata(br#"{"a":1,"b":2}"#))
9070                .unwrap()
9071                .content_hash
9072        );
9073    }
9074
9075    #[test]
9076    fn metadata_is_tamper_evident() {
9077        let card = br#"{"title":"x"}"#;
9078        let mut bytes = build_with_metadata(card);
9079        assert!(verify(&bytes).unwrap());
9080        // The card occupies [HEADER_LEN .. HEADER_LEN+card_len); flip a byte in it.
9081        bytes[HEADER_LEN + 2] ^= 0xff;
9082        assert!(
9083            !verify(&bytes).unwrap(),
9084            "tampering with the card must break verify()"
9085        );
9086    }
9087
9088    #[test]
9089    fn ranged_opens_do_not_fetch_metadata() {
9090        use crate::reader::{CountingReader, SliceReader};
9091        let card = vec![0xABu8; 512]; // distinctive and sizable
9092        let bytes = build_with_metadata(&card);
9093        let total = bytes.len() as u64;
9094
9095        // A full ranged open never loads the card and never reads its byte range.
9096        let r = CountingReader::new(SliceReader::new(&bytes));
9097        let rete = Rete::open_ranged(&r).unwrap();
9098        assert!(
9099            rete.metadata().is_none(),
9100            "open_ranged must not load the card"
9101        );
9102        assert!(r.requests() <= 4, "requests = {}", r.requests());
9103        assert!(
9104            r.bytes_read() <= total - card.len() as u64,
9105            "read {} of {} bytes; the {}-byte card must be skipped",
9106            r.bytes_read(),
9107            total,
9108            card.len()
9109        );
9110
9111        // Summary-only open likewise ignores the card and still summarizes.
9112        let rs = CountingReader::new(SliceReader::new(&bytes));
9113        let view = SummaryView::open_ranged(&rs).unwrap().unwrap();
9114        assert!(!view.summary.is_empty());
9115        assert!(rs.bytes_read() <= total - card.len() as u64);
9116    }
9117
9118    #[test]
9119    fn metadata_ranged_fetches_only_header_and_card() {
9120        use crate::reader::{CountingReader, SliceReader};
9121        // The CARD tier: fetch the self-description over a RangeReader touching
9122        // only the header + metadata range — never the dictionary/index/pyramid.
9123        let card = vec![0xCDu8; 384];
9124        let bytes = build_with_metadata(&card);
9125
9126        let r = CountingReader::new(SliceReader::new(&bytes));
9127        let got = read_metadata_ranged(&r).unwrap().unwrap();
9128        assert_eq!(got, card, "the card reads back verbatim");
9129        assert_eq!(r.requests(), 2, "exactly header + metadata ranges");
9130        assert_eq!(
9131            r.bytes_read(),
9132            HEADER_LEN as u64 + card.len() as u64,
9133            "no dictionary/index/pyramid bytes are touched"
9134        );
9135
9136        // A cardless file resolves to None after a single header read.
9137        let plain = build_image();
9138        let rp = CountingReader::new(SliceReader::new(&plain));
9139        assert!(read_metadata_ranged(&rp).unwrap().is_none());
9140        assert_eq!(rp.requests(), 1, "header only for a cardless file");
9141        assert_eq!(rp.bytes_read(), HEADER_LEN as u64);
9142    }
9143
9144    #[test]
9145    fn build_info_attaches_outside_the_hash() {
9146        let card = br#"{"title":"My Dataset"}"#;
9147        let base = build_with_metadata(card);
9148        let base_hash = Header::from_bytes(&base).unwrap().content_hash;
9149
9150        let info = br#"{"built_at":"2026-08-04T00:00:00Z","builder":"rete-cli 0.3.2"}"#;
9151        let with = attach_build_info(&base, info).unwrap();
9152
9153        // The section reads back verbatim, adjacent to the metadata.
9154        assert_eq!(read_build_info(&with).unwrap().unwrap(), info);
9155        let h = Header::from_bytes(&with).unwrap();
9156        assert_eq!(h.build_info_offset, HEADER_LEN as u64 + card.len() as u64);
9157        assert_eq!(h.build_info_len, info.len() as u64);
9158        assert_eq!(h.expected_file_len(), Some(with.len() as u64));
9159
9160        // OUTSIDE the hash: the content hash is unchanged and the file still
9161        // verifies — two builds of identical data stay hash-equal even though
9162        // their build-info differs.
9163        assert_eq!(h.content_hash, base_hash);
9164        assert!(verify(&with).unwrap());
9165        let other = attach_build_info(&base, br#"{"built_at":"1999-01-01T00:00:00Z"}"#).unwrap();
9166        assert_eq!(Header::from_bytes(&other).unwrap().content_hash, base_hash);
9167
9168        // The graph still opens and answers at its shifted offsets.
9169        let rete = Rete::open(&with).unwrap();
9170        assert_eq!(
9171            rete.query(Some("Bob"), Some("knows"), Some("Carol")).len(),
9172            1
9173        );
9174        assert_eq!(rete.metadata(), Some(card.as_slice()));
9175
9176        // Replacing an existing section works and removing it restores the
9177        // original image byte-for-byte (the strip-equality determinism proof).
9178        let replaced = attach_build_info(&with, br#"{"builder":"other"}"#).unwrap();
9179        assert_eq!(
9180            read_build_info(&replaced).unwrap().unwrap(),
9181            br#"{"builder":"other"}"#
9182        );
9183        assert!(verify(&replaced).unwrap());
9184        let stripped = attach_build_info(&replaced, &[]).unwrap();
9185        assert_eq!(stripped, base, "stripping build-info restores the image");
9186    }
9187
9188    #[test]
9189    fn build_info_attaches_on_a_metadata_free_file() {
9190        // No card: the section slots in right after the header.
9191        let base = build_image();
9192        let info = b"{\"builder\":\"x\"}";
9193        let with = attach_build_info(&base, info).unwrap();
9194        let h = Header::from_bytes(&with).unwrap();
9195        assert_eq!(h.build_info_offset, HEADER_LEN as u64);
9196        assert_eq!(h.metadata_len, 0);
9197        assert_eq!(read_build_info(&with).unwrap().unwrap(), info);
9198        assert!(verify(&with).unwrap());
9199        assert_eq!(
9200            h.content_hash,
9201            Header::from_bytes(&base).unwrap().content_hash
9202        );
9203        Rete::open(&with).unwrap();
9204    }
9205
9206    /// The contract a **streaming** rewriter needs: a plan is enough to rebuild
9207    /// the file from three spans — the header it hands back, the untouched
9208    /// `[HEADER_LEN, insert)` prefix, the new section, and the untouched
9209    /// `[tail_start, len)` tail — and the result is byte-identical to the
9210    /// in-memory splice. That is what lets a 17 GB file gain a build record
9211    /// with a 4 MiB buffer instead of 34 GB of RAM.
9212    #[test]
9213    fn a_plan_rebuilds_exactly_what_the_in_memory_splice_produces() {
9214        let card = br#"{"title":"My Dataset"}"#;
9215        let base = build_with_metadata(card);
9216        let base_hash = Header::from_bytes(&base).unwrap().content_hash;
9217        for start in [
9218            base.clone(),
9219            attach_build_info(&base, b"{\"a\":1}").unwrap(),
9220        ] {
9221            for info in [
9222                br#"{"builder":"rete-cli 0.3.2","query_costs":{}}"#.as_slice(), // longer
9223                b"{}".as_slice(),                                               // shorter
9224                b"".as_slice(),                                                 // removed
9225            ] {
9226                let plan = plan_build_info(&start, start.len() as u64, info.len() as u64).unwrap();
9227                let mut streamed = Vec::new();
9228                streamed.extend_from_slice(&plan.header);
9229                streamed.extend_from_slice(&start[HEADER_LEN..plan.insert as usize]);
9230                streamed.extend_from_slice(info);
9231                streamed.extend_from_slice(&start[plan.tail_start as usize..]);
9232                assert_eq!(streamed.len() as u64, plan.new_len);
9233                assert_eq!(
9234                    streamed,
9235                    attach_build_info(&start, info).unwrap(),
9236                    "a {}-byte section streamed must equal the same section spliced",
9237                    info.len()
9238                );
9239                // The point of the exercise: the file's identity is untouched.
9240                assert!(verify(&streamed).unwrap());
9241                let h = Header::from_bytes(&streamed).unwrap();
9242                assert_eq!(h.content_hash, base_hash);
9243                assert_eq!(h.expected_file_len(), Some(plan.new_len));
9244                let rete = Rete::open(&streamed).unwrap();
9245                assert_eq!(rete.metadata(), Some(card.as_slice()));
9246                assert_eq!(
9247                    rete.query(Some("Bob"), Some("knows"), Some("Carol")).len(),
9248                    1
9249                );
9250            }
9251        }
9252    }
9253
9254    /// A plan is refused rather than trusted when the header's own numbers put
9255    /// the section outside the file — the header is attacker-controlled input.
9256    #[test]
9257    fn a_plan_refuses_a_header_that_overruns_the_file() {
9258        let base = build_with_metadata(br#"{"title":"x"}"#);
9259        assert!(plan_build_info(&base, 8, 16).is_err());
9260        assert!(plan_build_info(&base[..HEADER_LEN], base.len() as u64, 16).is_ok());
9261    }
9262
9263    #[test]
9264    fn card_and_build_info_ranged_is_one_header_plus_one_range() {
9265        use crate::reader::{CountingReader, SliceReader};
9266        let card = vec![0xCDu8; 384];
9267        let info = vec![0xEFu8; 200];
9268        let bytes = attach_build_info(&build_with_metadata(&card), &info).unwrap();
9269
9270        // Both sections in ONE coalesced range after the header — the CARD
9271        // tier's 1 header + 1 range budget holds with build info included.
9272        let r = CountingReader::new(SliceReader::new(&bytes));
9273        let (m, b) = read_card_and_build_info_ranged(&r).unwrap();
9274        assert_eq!(m.unwrap(), card);
9275        assert_eq!(b.unwrap(), info);
9276        assert_eq!(r.requests(), 2, "header + one coalesced range");
9277        assert_eq!(
9278            r.bytes_read(),
9279            HEADER_LEN as u64 + card.len() as u64 + info.len() as u64
9280        );
9281
9282        // Card only: same two requests, no build info.
9283        let plain = build_with_metadata(&card);
9284        let rp = CountingReader::new(SliceReader::new(&plain));
9285        let (m, b) = read_card_and_build_info_ranged(&rp).unwrap();
9286        assert_eq!(m.unwrap(), card);
9287        assert!(b.is_none());
9288        assert_eq!(rp.requests(), 2);
9289        // Stated as the comparison every client depends on: carrying a build
9290        // record costs a reader NOTHING in requests. Any client that shows the
9291        // build record — the CLI, the browser card modal — can therefore do it
9292        // in the card's own budget, and does not need a second call.
9293        assert_eq!(
9294            r.requests(),
9295            rp.requests(),
9296            "build info must not cost an extra request",
9297        );
9298
9299        // Neither: the header read alone answers.
9300        let none = build_image();
9301        let rn = CountingReader::new(SliceReader::new(&none));
9302        let (m, b) = read_card_and_build_info_ranged(&rn).unwrap();
9303        assert!(m.is_none() && b.is_none());
9304        assert_eq!(rn.requests(), 1);
9305    }
9306
9307    #[test]
9308    fn schema_summary_groups_by_type() {
9309        let rt = RDF_TYPE;
9310        let bytes = build_from(&[
9311            ("Alice", rt, "Person"),
9312            ("Bob", rt, "Person"),
9313            ("NYC", rt, "City"),
9314            ("Alice", "knows", "Bob"),
9315            ("Alice", "livesIn", "NYC"),
9316            ("Alice", "name", "\"Alice\""),
9317        ]);
9318        let rete = Rete::open(&bytes).unwrap();
9319        let summary = schema_summary(&rete);
9320        // Expect class-level relations, rdf:type excluded.
9321        assert!(summary.contains(&("Person".into(), "knows".into(), "Person".into(), 1)));
9322        assert!(summary.contains(&("Person".into(), "livesIn".into(), "City".into(), 1)));
9323        assert!(summary.contains(&("Person".into(), "name".into(), "(literal)".into(), 1)));
9324        // No rdf:type relations in the summary.
9325        assert!(!summary.iter().any(|(_, p, _, _)| p == RDF_TYPE));
9326
9327        // Class populations: 2 People, 1 City, sorted by count desc.
9328        let classes = schema_classes(&rete);
9329        assert_eq!(
9330            classes,
9331            vec![("Person".into(), 2u32), ("City".into(), 1u32)]
9332        );
9333    }
9334
9335    fn build_from(triples: &[(&str, &str, &str)]) -> Vec<u8> {
9336        let mut db = DictionaryBuilder::new();
9337        for (s, p, o) in triples {
9338            db.observe(s, p, o);
9339        }
9340        let dict = db.build();
9341        let mut ib = GraphIndexBuilder::new();
9342        for (s, p, o) in triples {
9343            ib.push(dict.encode(s, p, o).unwrap());
9344        }
9345        write_file(&dict, &ib.build(), false, &[], 0)
9346    }
9347
9348    /// Build a file WITH a pyramid (so the schema pyramid + coherence axioms ship).
9349    fn build_with_pyramid(triples: &[(&str, &str, &str)]) -> Vec<u8> {
9350        let mut db = DictionaryBuilder::new();
9351        for (s, p, o) in triples {
9352            db.observe(s, p, o);
9353        }
9354        let dict = db.build();
9355        let encoded: Vec<_> = triples
9356            .iter()
9357            .map(|(s, p, o)| dict.encode(s, p, o).unwrap())
9358            .collect();
9359        let mut ib = GraphIndexBuilder::new();
9360        for t in &encoded {
9361            ib.push(*t);
9362        }
9363        let (meta, levels) = build_pyramid_meta(&dict, &encoded, DEFAULT_TILE_BUDGET);
9364        write_dataset(&dict, &ib.build(), &[], false, &meta, levels)
9365    }
9366
9367    #[test]
9368    fn tbox_coherence_flags_unsatisfiable_class_index_free() {
9369        use crate::reader::{CountingReader, SliceReader};
9370        let rt = RDF_TYPE;
9371        let sub = "<http://www.w3.org/2000/01/rdf-schema#subClassOf>";
9372        let disj = "<http://www.w3.org/2002/07/owl#disjointWith>";
9373        // C ⊑ D, C ⊑ E, D disjointWith E ⇒ C unsatisfiable — schema-only, with one
9374        // instance present just so the schema pyramid gets built.
9375        let bytes = build_with_pyramid(&[
9376            ("<http://ex/C>", sub, "<http://ex/D>"),
9377            ("<http://ex/C>", sub, "<http://ex/E>"),
9378            ("<http://ex/D>", disj, "<http://ex/E>"),
9379            ("<http://ex/x>", rt, "<http://ex/C>"),
9380        ]);
9381
9382        let r = CountingReader::new(SliceReader::new(&bytes));
9383        let view = SummaryView::open_ranged(&r).unwrap().unwrap();
9384        let points = view.tbox_coherence();
9385        assert!(
9386            points
9387                .iter()
9388                .any(|i| i.kind == "unsatisfiable-class" && i.detail.contains("http://ex/C>")),
9389            "expected C unsatisfiable from the schema alone, got {points:?}"
9390        );
9391
9392        // Proven index-free: bytes read never reach the (root_dir) index section,
9393        // mirroring `schema_pyramid_round_trips_through_file_index_free`.
9394        let header = Header::from_bytes(&bytes[..HEADER_LEN]).unwrap();
9395        assert!(
9396            r.bytes_read() <= bytes.len() as u64 - header.root_dir_len,
9397            "tbox_coherence must not read the triple index"
9398        );
9399    }
9400
9401    #[test]
9402    fn schema_coherence_reads_only_the_schema_block() {
9403        use crate::reader::{CountingReader, SliceReader};
9404        let rt = RDF_TYPE;
9405        let sub = "<http://www.w3.org/2000/01/rdf-schema#subClassOf>";
9406        let disj = "<http://www.w3.org/2002/07/owl#disjointWith>";
9407        // 500 instances with unique literals → a sizable dictionary + community
9408        // summary, so a whole-pyramid-meta read would be large; the schema block
9409        // (bounded by the tiny ontology) stays small.
9410        let mut triples: Vec<(String, String, String)> = vec![
9411            ("<http://ex/C>".into(), sub.into(), "<http://ex/D>".into()),
9412            ("<http://ex/C>".into(), sub.into(), "<http://ex/E>".into()),
9413            ("<http://ex/D>".into(), disj.into(), "<http://ex/E>".into()),
9414        ];
9415        for i in 0..500 {
9416            let s = format!("<http://ex/x{i}>");
9417            triples.push((s.clone(), rt.into(), "<http://ex/C>".into()));
9418            triples.push((
9419                s,
9420                "<http://ex/label>".into(),
9421                format!("\"unique label {i}\""),
9422            ));
9423        }
9424        let trefs: Vec<(&str, &str, &str)> = triples
9425            .iter()
9426            .map(|(s, p, o)| (s.as_str(), p.as_str(), o.as_str()))
9427            .collect();
9428        let bytes = build_with_pyramid(&trefs);
9429
9430        let header = Header::from_bytes(&bytes[..HEADER_LEN]).unwrap();
9431        assert!(
9432            header.schema_meta_len > 0,
9433            "the writer recorded a schema-block length"
9434        );
9435        assert!(
9436            (header.schema_meta_len as u64) < header.pyramid_meta_len,
9437            "schema block ({}) should be far smaller than the whole pyramid-meta ({})",
9438            header.schema_meta_len,
9439            header.pyramid_meta_len
9440        );
9441
9442        let r = CountingReader::new(SliceReader::new(&bytes));
9443        let points = read_schema_coherence_ranged(&r).unwrap().unwrap();
9444        assert!(points.iter().any(|i| i.kind == "unsatisfiable-class"));
9445        // It read only the header + the schema block — not the summary or dictionary.
9446        assert!(
9447            r.bytes_read() <= HEADER_LEN as u64 + header.schema_meta_len as u64,
9448            "read {} bytes; expected <= header + schema block ({})",
9449            r.bytes_read(),
9450            HEADER_LEN as u64 + header.schema_meta_len as u64
9451        );
9452    }
9453
9454    #[test]
9455    fn tbox_coherence_clean_schema_is_coherent() {
9456        use crate::reader::SliceReader;
9457        let rt = RDF_TYPE;
9458        let sub = "<http://www.w3.org/2000/01/rdf-schema#subClassOf>";
9459        let bytes = build_with_pyramid(&[
9460            ("<http://ex/Dog>", sub, "<http://ex/Animal>"),
9461            ("<http://ex/x>", rt, "<http://ex/Dog>"),
9462        ]);
9463        let view = SummaryView::open_ranged(&SliceReader::new(&bytes))
9464            .unwrap()
9465            .unwrap();
9466        assert!(view.tbox_is_coherent(), "a plain hierarchy is coherent");
9467    }
9468
9469    /// A quad-index file must be **refused** by every reader that predates it —
9470    /// an old reader ignores flag bit 3 and the unknown kind 10, so it would
9471    /// otherwise open the file as one without named graphs and answer `GRAPH`
9472    /// queries with nothing. The seal makes each of their open paths fail on the
9473    /// index container's section count, before any payload is read; their
9474    /// `verify` still passes, because the part list it hashes is unchanged.
9475    /// (Checked against real 0.3.x binaries too: QUAD-INDEX-DESIGN.md §9.)
9476    #[test]
9477    fn readers_before_the_quad_index_refuse_its_files() {
9478        for perms in [PermSet::ALL, PermSet::CORE] {
9479            let quads: Vec<crate::ingest::RawQuad> = (0..50)
9480                .map(|i| {
9481                    (
9482                        format!("<http://ex/s{i}>"),
9483                        "<http://ex/p>".to_string(),
9484                        format!("\"{i}\""),
9485                        Some(format!("<http://ex/g{}>", i % 7)),
9486                    )
9487                })
9488                .collect();
9489            let (bytes, _) = crate::ingest::assemble_dataset_with_layout(
9490                quads,
9491                false,
9492                false,
9493                None,
9494                crate::PyramidAlgo::Louvain,
9495                perms,
9496                crate::ingest::GraphLayout::Quad,
9497                |_, _| Vec::<u8>::new(),
9498            );
9499            let h = Header::from_bytes(&bytes).unwrap();
9500            assert!(h.has_quad_index());
9501            let root =
9502                &bytes[h.root_dir_offset as usize..(h.root_dir_offset + h.root_dir_len) as usize];
9503            // 0.3.4–0.3.7, resident (`Rete::open`, `open_ranged`): count != mask.
9504            assert!(decode_index_container(root, h.block_codec, h.perms).is_err());
9505            assert!(
9506                container_section_payload_ranges(root, h.root_dir_offset, h.perms.len()).is_err()
9507            );
9508            // 0.3.0–0.3.3 expected exactly six sections, whatever the mask.
9509            assert_ne!(decode_container(root, CODEC_NONE).unwrap().len(), NUM_PERMS);
9510            assert_ne!(
9511                decode_container(root, CODEC_NONE).unwrap().len(),
9512                h.perms.len()
9513            );
9514            // Every lazy / remote open (`open_ranged_lazy`, `query_ranged`).
9515            let r = crate::reader::SliceReader::new(&bytes);
9516            for expect in [h.perms.len(), NUM_PERMS] {
9517                assert!(locate_container_section_ranged(
9518                    &r,
9519                    h.root_dir_offset,
9520                    h.root_dir_len,
9521                    0,
9522                    expect as u64
9523                )
9524                .is_err());
9525            }
9526            // Their `verify` hashes the same parts the writer hashed.
9527            assert_eq!(hash_of_sections(&h, &bytes).unwrap(), h.content_hash);
9528            // And a current reader reads the file.
9529            let rete = Rete::open(&bytes).unwrap();
9530            assert_eq!(rete.named_graph_count(), 7);
9531            assert_eq!(rete.dump(Some("<http://ex/g3>")).len(), 7);
9532        }
9533    }
9534
9535    #[test]
9536    fn named_graphs_round_trip() {
9537        // One shared dictionary; default graph + a named graph "g1".
9538        let all = [
9539            ("Alice", "knows", "Bob"), // default
9540            ("Bob", "age", "30"),      // named g1
9541        ];
9542        let mut db = DictionaryBuilder::new();
9543        for (s, p, o) in all {
9544            db.observe(s, p, o);
9545        }
9546        let dict = db.build();
9547
9548        let mut def = GraphIndexBuilder::new();
9549        def.push(dict.encode("Alice", "knows", "Bob").unwrap());
9550        let mut g1 = GraphIndexBuilder::new();
9551        g1.push(dict.encode("Bob", "age", "30").unwrap());
9552
9553        let named = vec![("http://ex/g1".to_string(), g1.build())];
9554        let bytes = write_dataset(&dict, &def.build(), &named, true, &[], 0);
9555
9556        assert!(verify(&bytes).unwrap());
9557        let rete = Rete::open(&bytes).unwrap();
9558        assert_eq!(rete.graph_names(), vec!["http://ex/g1"]);
9559        // The named graph contains Bob age 30, not the default-graph triple.
9560        let gi = rete.graph_index("http://ex/g1").unwrap();
9561        assert_eq!(gi.triple_count(), 1);
9562        assert!(rete.graph_index("http://ex/missing").is_none());
9563
9564        // quad_count counts ALL quads — default graph + named graphs (1 + 1),
9565        // not just the default index (which would report 1).
9566        assert_eq!(rete.header().quad_count, 2);
9567
9568        // Default-graph query path is unchanged.
9569        assert_eq!(rete.query(Some("Alice"), None, None).len(), 1);
9570
9571        // dump() round-trips each graph back to terms.
9572        assert_eq!(
9573            rete.dump(None),
9574            vec![("Alice".into(), "knows".into(), "Bob".into())]
9575        );
9576        assert_eq!(
9577            rete.dump(Some("http://ex/g1")),
9578            vec![("Bob".into(), "age".into(), "30".into())]
9579        );
9580    }
9581
9582    /// A named graph whose container exceeds [`NAMED_GRAPH_RESIDENT_MAX`]
9583    /// opens through [`open_index_container_lazy`] — the default graph's own
9584    /// machinery, at the container's offset. Drive that helper directly over a
9585    /// named graph's container range (the threshold itself is too large to
9586    /// cross cheaply in a unit test) and require identical matches.
9587    #[test]
9588    fn named_graph_container_opens_tile_lazily_like_the_root() {
9589        let mut db = DictionaryBuilder::new();
9590        let node = |n: u32| format!("<http://ex/n{n}>");
9591        let p = "<http://ex/p>".to_string();
9592        db.observe(&node(0), &p, &node(1));
9593        for i in 0..500u32 {
9594            db.observe(&node(1000 + i), &p, &node(2000 + i));
9595        }
9596        let dict = db.build();
9597        let mut def = GraphIndexBuilder::new();
9598        def.push(dict.encode(&node(0), &p, &node(1)).unwrap());
9599        let mut g = GraphIndexBuilder::new().with_tile_budget(256); // many tiles
9600        for i in 0..500u32 {
9601            g.push(dict.encode(&node(1000 + i), &p, &node(2000 + i)).unwrap());
9602        }
9603        let named = vec![("<http://ex/g>".to_string(), g.build())];
9604        let bytes = write_dataset(&dict, &def.build(), &named, true, &[], 0);
9605        let rete = Rete::open(&bytes).unwrap();
9606        let header = rete.header();
9607
9608        // The single entry's container range: count varint, iri_len varint,
9609        // iri, container_len varint, then the container.
9610        let soff = header.named_graphs_offset as usize;
9611        let send = soff + header.named_graphs_len as usize;
9612        let sec = &bytes[soff..send];
9613        let (n, mut pos) = read_uvarint(sec).unwrap();
9614        assert_eq!(n, 1);
9615        let (ilen, u1) = read_uvarint(&sec[pos..]).unwrap();
9616        pos += u1 + ilen as usize;
9617        let (clen, u2) = read_uvarint(&sec[pos..]).unwrap();
9618        pos += u2;
9619        let container = ByteRange {
9620            offset: (soff + pos) as u64,
9621            len: clen,
9622        };
9623
9624        use crate::reader::SliceReader;
9625        let leaked: &'static [u8] = Box::leak(bytes.clone().into_boxed_slice());
9626        let reader: std::sync::Arc<dyn RangeReader + Send + Sync> =
9627            std::sync::Arc::new(SliceReader::new(leaked));
9628        let (lazy_idx, _, _) = open_index_container_lazy(
9629            &reader,
9630            container,
9631            header.block_codec,
9632            header.has_tile_synopsis(),
9633            1,
9634            header.perms,
9635            header.perms.len(),
9636        )
9637        .unwrap();
9638        let mut want = rete
9639            .graph_index("<http://ex/g>")
9640            .unwrap()
9641            .match_pattern((None, None, None));
9642        let mut got = lazy_idx.match_pattern((None, None, None));
9643        want.sort_unstable();
9644        got.sort_unstable();
9645        assert_eq!(got.len(), 500);
9646        assert_eq!(got, want);
9647        assert!(!lazy_idx.load_incomplete());
9648    }
9649
9650    /// `dump_iter` must agree with `dump` exactly — and must be *lazy*: taking
9651    /// the first triple of a lazily range-read file may not drag the whole
9652    /// index across the reader, or the JS client's streaming quad cursor (which
9653    /// is this iterator, suspended between wasm calls) would be a materializing
9654    /// dump wearing a cursor's clothes.
9655    #[test]
9656    fn dump_iter_matches_dump_and_stops_early() {
9657        use crate::reader::{CountingReader, SliceReader};
9658
9659        // Enough triples over tiny tiles to force many tiles per permutation.
9660        let triples: Vec<(String, String, String)> = (0..2000u32)
9661            .map(|i| {
9662                (
9663                    format!("<http://ex/s/{i:04}>"),
9664                    "<http://ex/p>".to_string(),
9665                    format!("<http://ex/o/{i:04}>"),
9666                )
9667            })
9668            .collect();
9669        let mut db = DictionaryBuilder::new();
9670        for (s, p, o) in &triples {
9671            db.observe(s, p, o);
9672        }
9673        let dict = db.build();
9674        let mut ib = GraphIndexBuilder::new().with_tile_budget(256);
9675        for (s, p, o) in &triples {
9676            ib.push(dict.encode(s, p, o).unwrap());
9677        }
9678        let bytes = write_file(&dict, &ib.build(), false, &[], 0);
9679
9680        // Same triples, same order, as the eager dump.
9681        let rete = Rete::open(&bytes).unwrap();
9682        assert_eq!(rete.dump_iter(None).collect::<Vec<_>>(), rete.dump(None));
9683        // A missing named graph yields nothing rather than the default graph.
9684        assert_eq!(rete.dump_iter(Some("http://ex/nope")).count(), 0);
9685
9686        // Lazily: taking one triple must cost far fewer bytes than draining.
9687        let read_bytes = |take: usize| -> u64 {
9688            let leaked: &'static [u8] = Box::leak(bytes.clone().into_boxed_slice());
9689            let reader = std::sync::Arc::new(CountingReader::new(SliceReader::new(leaked)));
9690            let rete = Rete::open_ranged_lazy_with(reader.clone(), HEADER_LEN as u64).unwrap();
9691            let before = reader.bytes_read();
9692            let n = rete.dump_iter(None).take(take).count();
9693            assert_eq!(n, take.min(triples.len()));
9694            assert!(!rete.index_incomplete());
9695            reader.bytes_read() - before
9696        };
9697        let first_only = read_bytes(1);
9698        let everything = read_bytes(triples.len());
9699        // Always true: stopping after one triple cannot cost a full scan.
9700        assert!(
9701            first_only < everything,
9702            "dump_iter is not lazy: one triple read {first_only} B, a full scan {everything} B"
9703        );
9704        // The STRONG claim — under half — only holds once the payload dwarfs the
9705        // fixed open cost. This fixture is ~32 KB, and with `compression` off the
9706        // dictionary is essentially the whole file, so opening it at all already
9707        // pays most of `everything` (21 KB of 33 KB) and the ratio says nothing
9708        // about laziness. Asserting it unconditionally is how this test failed on
9709        // `--no-default-features` while passing everywhere else.
9710        #[cfg(feature = "compression")]
9711        assert!(
9712            first_only * 2 < everything,
9713            "dump_iter is not lazy: one triple read {first_only} B of the {everything} B a full scan reads"
9714        );
9715
9716        // A FRESH open: these assertions fault tiles and the dictionary in,
9717        // and the laziness measurement above counts bytes on a COLD reader —
9718        // sharing one `rete` makes first_only and everything both ~0, so the
9719        // comparison silently stops testing anything.
9720        let fresh = Rete::open(&bytes).unwrap();
9721        // `dump_batch` is the resumable form the language clients pull through.
9722        // Driven to exhaustion it must reproduce the eager dump EXACTLY — same
9723        // triples, same order — for any batch size, including one so small that
9724        // every call ends on a subject boundary before reaching it.
9725        let drain = |max_quads: usize| {
9726            let (mut out, mut cursor, mut calls) = (Vec::new(), 0u32, 0);
9727            loop {
9728                let (batch, next, done) = fresh.dump_batch(None, cursor, max_quads);
9729                out.extend(batch);
9730                cursor = next;
9731                calls += 1;
9732                assert!(calls < 10_000, "cursor failed to advance");
9733                if done {
9734                    break out;
9735                }
9736            }
9737        };
9738        let eager = fresh.dump(None);
9739        for max in [1usize, 7, 128, 100_000] {
9740            assert_eq!(drain(max), eager, "batch size {max} changed the dump");
9741        }
9742        // An unknown graph is an empty, already-finished dump — not an error and
9743        // not the default graph leaking through.
9744        let (t, _, done) = fresh.dump_batch(Some("http://ex/nope"), 0, 16);
9745        assert!(t.is_empty() && done);
9746    }
9747
9748    /// `query_batch` is `dump_batch` generalized to a PATTERN — the primitive the
9749    /// Java client's RDF4J cursor pulls through, one bounded batch per wasm call.
9750    /// Three things have to hold or the cursor is a materializing scan wearing a
9751    /// cursor's clothes: driven to exhaustion it must reproduce the eager
9752    /// `query_in_graph` for every pattern shape and every batch size; taking one
9753    /// row must cost far fewer bytes than draining; and every call must either
9754    /// yield a row or say `done`, so the caller can never spin.
9755    #[test]
9756    fn query_batch_matches_query_in_graph_and_stops_early() {
9757        use crate::reader::{CountingReader, SliceReader};
9758
9759        // Enough triples over tiny tiles that a scan spans many tiles, with two
9760        // predicates and a repeated subject so bound patterns have real fanout.
9761        let triples: Vec<(String, String, String)> = (0..2000u32)
9762            .map(|i| {
9763                (
9764                    format!("<http://ex/s/{:04}>", i / 2),
9765                    format!("<http://ex/p/{}>", i % 2),
9766                    format!("<http://ex/o/{i:04}>"),
9767                )
9768            })
9769            .collect();
9770        let mut db = DictionaryBuilder::new();
9771        for (s, p, o) in &triples {
9772            db.observe(s, p, o);
9773        }
9774        let dict = db.build();
9775        let mut ib = GraphIndexBuilder::new().with_tile_budget(256);
9776        for (s, p, o) in &triples {
9777            ib.push(dict.encode(s, p, o).unwrap());
9778        }
9779        let bytes = write_file(&dict, &ib.build(), false, &[], 0);
9780        let rete = Rete::open(&bytes).unwrap();
9781
9782        let drain = |s: Option<&str>, p: Option<&str>, o: Option<&str>, max: usize| {
9783            let (mut out, mut cursor, mut calls) = (Vec::new(), 0u64, 0);
9784            loop {
9785                let (batch, next, done) = rete.query_batch(None, s, p, o, cursor, max);
9786                // The contract the client relies on: rows, or done. Never
9787                // "nothing yet, call again" — that is an unbounded spin.
9788                assert!(!batch.is_empty() || done, "empty non-final batch");
9789                out.extend(batch);
9790                cursor = next;
9791                calls += 1;
9792                assert!(calls < 20_000, "cursor failed to advance");
9793                if done {
9794                    break out;
9795                }
9796            }
9797        };
9798        let shapes: [(Option<&str>, Option<&str>, Option<&str>); 6] = [
9799            (None, None, None),
9800            (Some("<http://ex/s/0007>"), None, None),
9801            (None, Some("<http://ex/p/1>"), None),
9802            (None, None, Some("<http://ex/o/0042>")),
9803            (Some("<http://ex/s/0007>"), Some("<http://ex/p/0>"), None),
9804            (
9805                Some("<http://ex/s/0007>"),
9806                Some("<http://ex/p/0>"),
9807                Some("<http://ex/o/0014>"),
9808            ),
9809        ];
9810        for (s, p, o) in shapes {
9811            let eager = rete.query_in_graph(None, s, p, o);
9812            assert!(
9813                !eager.is_empty(),
9814                "fixture shape {s:?} {p:?} {o:?} matches nothing"
9815            );
9816            // The lazy pull form agrees exactly, order included.
9817            assert_eq!(
9818                rete.query_iter(None, s, p, o).collect::<Vec<_>>(),
9819                eager,
9820                "query_iter disagrees for {s:?} {p:?} {o:?}"
9821            );
9822            for max in [1usize, 3, 64, 100_000] {
9823                let mut got = drain(s, p, o, max);
9824                let mut want = eager.clone();
9825                // A batch is re-sorted canonically per batch, not globally, so
9826                // a bound pattern's batches can arrive in permutation order.
9827                got.sort();
9828                want.sort();
9829                assert_eq!(got, want, "batch size {max} changed {s:?} {p:?} {o:?}");
9830            }
9831            // The fully unbound scan routes to SPO, so batching is order-exact.
9832            if s.is_none() && p.is_none() && o.is_none() {
9833                assert_eq!(drain(s, p, o, 7), eager, "unbound batching reordered rows");
9834            }
9835        }
9836
9837        // A bound term the dictionary has never seen, and an unknown graph, are
9838        // both finished-and-empty rather than errors or a leak of everything.
9839        let (t, _, done) = rete.query_batch(None, Some("<http://ex/nope>"), None, None, 0, 16);
9840        assert!(t.is_empty() && done);
9841        let (t, _, done) = rete.query_batch(Some("http://ex/nope"), None, None, None, 0, 16);
9842        assert!(t.is_empty() && done);
9843        assert_eq!(
9844            rete.query_iter(None, Some("<http://ex/nope>"), None, None)
9845                .count(),
9846            0
9847        );
9848
9849        // Lazily: one row must not drag the whole scan across the reader. This
9850        // is the property that makes `LIMIT 1` bounded on a 48 GiB file.
9851        let read_bytes = |take: usize| -> u64 {
9852            let leaked: &'static [u8] = Box::leak(bytes.clone().into_boxed_slice());
9853            let reader = std::sync::Arc::new(CountingReader::new(SliceReader::new(leaked)));
9854            let rete = Rete::open_ranged_lazy_with(reader.clone(), HEADER_LEN as u64).unwrap();
9855            let before = reader.bytes_read();
9856            let (mut n, mut cursor) = (0usize, 0u64);
9857            loop {
9858                let (batch, next, done) = rete.query_batch(None, None, None, None, cursor, 1);
9859                n += batch.len();
9860                cursor = next;
9861                if done || n >= take {
9862                    break;
9863                }
9864            }
9865            assert!(n >= take.min(triples.len()));
9866            assert!(!rete.index_incomplete());
9867            reader.bytes_read() - before
9868        };
9869        let first_only = read_bytes(1);
9870        let everything = read_bytes(triples.len());
9871        assert!(
9872            first_only < everything,
9873            "query_batch is not lazy: one row read {first_only} B, a full scan {everything} B"
9874        );
9875        #[cfg(feature = "compression")]
9876        assert!(
9877            first_only * 2 < everything,
9878            "query_batch is not lazy: one row read {first_only} B of the {everything} B a full scan reads"
9879        );
9880    }
9881
9882    /// A batched cursor over a QUADS file must stay inside the graph it was
9883    /// given: the default graph and each named graph resume independently.
9884    #[test]
9885    fn query_batch_is_graph_scoped() {
9886        let mut db = DictionaryBuilder::new();
9887        for (s, p, o) in [
9888            ("Alice", "knows", "Bob"),
9889            ("Alice", "knows", "Carol"),
9890            ("Alice", "knows", "Dave"),
9891            ("Erin", "knows", "Frank"),
9892            ("Alice", "knows", "Zoe"),
9893        ] {
9894            db.observe(s, p, o);
9895        }
9896        let dict = db.build();
9897        let mut def = GraphIndexBuilder::new();
9898        def.push(dict.encode("Alice", "knows", "Bob").unwrap());
9899        def.push(dict.encode("Alice", "knows", "Carol").unwrap());
9900        let mut g1 = GraphIndexBuilder::new();
9901        g1.push(dict.encode("Alice", "knows", "Dave").unwrap());
9902        g1.push(dict.encode("Erin", "knows", "Frank").unwrap());
9903        let mut g2 = GraphIndexBuilder::new();
9904        g2.push(dict.encode("Alice", "knows", "Zoe").unwrap());
9905        let named = vec![
9906            ("http://ex/g1".to_string(), g1.build()),
9907            ("http://ex/g2".to_string(), g2.build()),
9908        ];
9909        let bytes = write_dataset(&dict, &def.build(), &named, true, &[], 0);
9910        let rete = Rete::open(&bytes).unwrap();
9911        for graph in [None, Some("http://ex/g1"), Some("http://ex/g2")] {
9912            let eager = rete.query_in_graph(graph, None, None, None);
9913            let (mut out, mut cursor) = (Vec::new(), 0u64);
9914            loop {
9915                let (batch, next, done) = rete.query_batch(graph, None, None, None, cursor, 1);
9916                out.extend(batch);
9917                cursor = next;
9918                if done {
9919                    break;
9920                }
9921            }
9922            assert_eq!(out, eager, "graph {graph:?} batched differently");
9923        }
9924    }
9925
9926    #[test]
9927    fn query_in_graph_is_graph_scoped() {
9928        // Default graph: Alice knows Bob, Alice knows Carol.
9929        // Named g1: Alice knows Dave (same predicate, different graph).
9930        let mut db = DictionaryBuilder::new();
9931        for (s, p, o) in [
9932            ("Alice", "knows", "Bob"),
9933            ("Alice", "knows", "Carol"),
9934            ("Alice", "knows", "Dave"),
9935        ] {
9936            db.observe(s, p, o);
9937        }
9938        let dict = db.build();
9939
9940        let mut def = GraphIndexBuilder::new();
9941        def.push(dict.encode("Alice", "knows", "Bob").unwrap());
9942        def.push(dict.encode("Alice", "knows", "Carol").unwrap());
9943        let mut g1 = GraphIndexBuilder::new();
9944        g1.push(dict.encode("Alice", "knows", "Dave").unwrap());
9945
9946        let named = vec![("http://ex/g1".to_string(), g1.build())];
9947        let bytes = write_dataset(&dict, &def.build(), &named, true, &[], 0);
9948        let rete = Rete::open(&bytes).unwrap();
9949
9950        // Default graph only: the two default-graph objects, not Dave.
9951        let mut def_objs: Vec<String> = rete
9952            .query_in_graph(None, Some("Alice"), Some("knows"), None)
9953            .into_iter()
9954            .map(|(_, _, o)| o)
9955            .collect();
9956        def_objs.sort();
9957        assert_eq!(def_objs, vec!["Bob".to_string(), "Carol".to_string()]);
9958
9959        // Named graph only: just Dave.
9960        assert_eq!(
9961            rete.query_in_graph(Some("http://ex/g1"), Some("Alice"), None, None),
9962            vec![("Alice".into(), "knows".into(), "Dave".into())]
9963        );
9964
9965        // A wildcard-everything scan is scoped to its graph.
9966        assert_eq!(rete.query_in_graph(None, None, None, None).len(), 2);
9967        assert_eq!(
9968            rete.query_in_graph(Some("http://ex/g1"), None, None, None)
9969                .len(),
9970            1
9971        );
9972
9973        // An unknown graph IRI is empty, not an error.
9974        assert!(rete
9975            .query_in_graph(Some("http://ex/missing"), None, None, None)
9976            .is_empty());
9977    }
9978
9979    /// A **bounded** `query_in_graph` — and a dump of one named graph — must not
9980    /// fault the whole dictionary on a lazily range-read file.
9981    ///
9982    /// Both used to call `Dictionary::prefetch_all`, so the graph-scoped pattern
9983    /// primitive an RDF4J `Sail` calls once per named graph paid for every term
9984    /// in the file to answer a handful of rows. Measured on `cordis.rete`
9985    /// (801 MB, a 417 MB dictionary) before the fix: a 15-row answer read
9986    /// 416 MB and peaked at 2.5 GB RSS. The fixture below is the same shape in
9987    /// miniature — long literals so the dictionary dominates the file — and the
9988    /// assertion is the invariant that regressed: bytes read must stay well
9989    /// under the dictionary section. Lazy answers are compared against eager
9990    /// throughout: cheaper is only interesting if it is also correct.
9991    #[test]
9992    fn bounded_query_in_graph_does_not_fault_the_whole_dictionary() {
9993        use crate::reader::{CountingReader, SliceReader};
9994
9995        // Long, distinct, poorly-compressible object literals: the dictionary is
9996        // the payload here, exactly as on a graph that stores abstracts or
9997        // embedded media. Repeating one filler string would not do — front
9998        // coding plus zstd shrink 2000 near-identical literals to ~6 KB, which
9999        // is smaller than the index and makes the measurement meaningless.
10000        let payload = |i: u32| -> String {
10001            let mut s = String::with_capacity(800);
10002            let mut x = i as u64 ^ 0x9e37_79b9_7f4a_7c15;
10003            while s.len() < 800 {
10004                x = x
10005                    .wrapping_mul(6364136223846793005)
10006                    .wrapping_add(1442695040888963407);
10007                s.push_str(&format!("{:016x}", x ^ (x >> 29)));
10008            }
10009            s
10010        };
10011        // Half the quads go to the named graph, and the two halves are kept
10012        // lexically apart (`a…` / `b…`) so each lands in its own dictionary
10013        // chunks — the real layout on a quads file whose graphs use different
10014        // IRI namespaces, e.g. `cordis.rete`. Interleaving them instead would
10015        // put a term of both graphs in every chunk, and then no per-graph
10016        // access pattern could avoid faulting the whole dictionary.
10017        let quads: Vec<(String, String, String, bool)> = (0..2000u32)
10018            .map(|i| {
10019                let named = i % 2 == 0;
10020                let ns = if named { 'a' } else { 'b' };
10021                (
10022                    format!("<http://ex/{ns}/s/{i:04}>"),
10023                    "<http://ex/abstract>".to_string(),
10024                    format!("\"{ns}{} number {i}\"", payload(i)),
10025                    named,
10026                )
10027            })
10028            .collect();
10029        let mut db = DictionaryBuilder::new();
10030        for (s, p, o, _) in &quads {
10031            db.observe(s, p, o);
10032        }
10033        let dict = db.build();
10034        let mut def = GraphIndexBuilder::new().with_tile_budget(256);
10035        let mut g1 = GraphIndexBuilder::new().with_tile_budget(256);
10036        for (s, p, o, named) in &quads {
10037            let t = dict.encode(s, p, o).unwrap();
10038            if *named {
10039                g1.push(t);
10040            } else {
10041                def.push(t);
10042            }
10043        }
10044        let named = vec![("http://ex/g1".to_string(), g1.build())];
10045        let bytes = write_dataset(&dict, &def.build(), &named, true, &[], 0);
10046
10047        let eager = Rete::open(&bytes).unwrap();
10048        let dict_len = eager.header().dictionary_len;
10049
10050        // `probe` runs one closure against a COLD lazy open and reports the
10051        // bytes it faulted beyond the open itself.
10052        let leaked: &'static [u8] = Box::leak(bytes.clone().into_boxed_slice());
10053        let probe = |f: &dyn Fn(&Rete) -> Vec<TermTriple>| -> (Vec<TermTriple>, u64) {
10054            let reader = std::sync::Arc::new(CountingReader::new(SliceReader::new(leaked)));
10055            let lazy = Rete::open_ranged_lazy_with(reader.clone(), HEADER_LEN as u64).unwrap();
10056            let before = reader.bytes_read();
10057            let out = f(&lazy);
10058            assert!(!lazy.index_incomplete(), "a chunk fetch failed");
10059            (out, reader.bytes_read() - before)
10060        };
10061
10062        // 1. One bound subject in the named graph.
10063        let want = eager.query_in_graph(
10064            Some("http://ex/g1"),
10065            Some("<http://ex/a/s/0100>"),
10066            None,
10067            None,
10068        );
10069        assert_eq!(want.len(), 1, "fixture: expected exactly one match");
10070        let (got, pulled) = probe(&|r| {
10071            r.query_in_graph(
10072                Some("http://ex/g1"),
10073                Some("<http://ex/a/s/0100>"),
10074                None,
10075                None,
10076            )
10077        });
10078        assert_eq!(got, want, "lazy query_in_graph disagrees with eager");
10079        assert!(
10080            pulled * 4 < dict_len,
10081            "a 1-row query_in_graph faulted {pulled} B of a {dict_len} B dictionary — \
10082             it is prefetching the whole thing again"
10083        );
10084
10085        // 2. A graph with NOTHING to resolve must not fault the dictionary.
10086        //    Both shapes below used to pull every chunk before returning an
10087        //    empty answer, and `rete export` hits the second one first on a
10088        //    quads file whose default graph is empty — on `cordis.rete` that
10089        //    was 415 MB read and 2.5 GB of peak RSS to emit zero triples. What
10090        //    is left is the named-graph directory walk it takes to learn the
10091        //    graph is absent at all; that is index, not dictionary.
10092        let (empty, q_pulled) =
10093            probe(&|r| r.query_in_graph(Some("http://ex/nope"), None, None, None));
10094        assert!(empty.is_empty());
10095        let (_, d_pulled) = probe(&|r| {
10096            let mut n = 0usize;
10097            r.dump_each(Some("http://ex/nope"), |_, _, _| n += 1);
10098            assert_eq!(n, 0, "an absent graph yielded triples");
10099            Vec::new()
10100        });
10101        assert_eq!(
10102            q_pulled, d_pulled,
10103            "query_in_graph and dump_each disagree on what an absent graph costs"
10104        );
10105        assert!(
10106            q_pulled * 4 < dict_len,
10107            "an absent graph faulted {q_pulled} B of a {dict_len} B dictionary"
10108        );
10109
10110        // 3. Dumping ONE named graph costs that graph, not the whole file: the
10111        //    default graph's half of the dictionary must stay unfaulted.
10112        let want: Vec<TermTriple> = eager.dump(Some("http://ex/g1"));
10113        assert_eq!(want.len(), 1000);
10114        let (got, pulled) = probe(&|r| {
10115            let mut out = Vec::new();
10116            r.dump_each(Some("http://ex/g1"), |s, p, o| {
10117                out.push((s.to_string(), p.to_string(), o.to_string()))
10118            });
10119            out
10120        });
10121        assert_eq!(got, want, "lazy dump_each disagrees with eager dump");
10122        assert!(
10123            pulled * 3 < dict_len * 2,
10124            "dumping half the graph faulted {pulled} B of a {dict_len} B dictionary — \
10125             the other graph's chunks are being pulled in too"
10126        );
10127    }
10128
10129    /// Build a multi-chunk single-graph ranged file image for the eviction
10130    /// tests: 4000 distinct ~70-byte object literals overflow the 64 KiB
10131    /// dictionary chunk budget into several chunks, and one multi-KiB literal
10132    /// forces a large run/chunk boundary — the shape a tiny budget must survive.
10133    #[cfg(test)]
10134    fn build_multichunk_image() -> Vec<u8> {
10135        use crate::index::PermSet;
10136        let mut db = DictionaryBuilder::new();
10137        let mut triples: Vec<(String, String, String)> = (0..4000u32)
10138            .map(|i| {
10139                (
10140                    format!("<http://ex/s/{i:07}>"),
10141                    format!("<http://ex/p/{}>", i % 8),
10142                    format!("\"object literal number {i:07} with padding text to widen the term\""),
10143                )
10144            })
10145            .collect();
10146        triples.push((
10147            "<http://ex/s/big>".to_string(),
10148            "<http://ex/p/big>".to_string(),
10149            format!("\"{}\"", "Z".repeat(8000)),
10150        ));
10151        for (s, p, o) in &triples {
10152            db.observe(s, p, o);
10153        }
10154        let dict = db.build();
10155        let mut idx = GraphIndexBuilder::new()
10156            .with_tile_budget(64)
10157            .with_perms(PermSet::ALL);
10158        for (s, p, o) in &triples {
10159            idx.push(dict.encode(s, p, o).unwrap());
10160        }
10161        write_dataset(&dict, &idx.build(), &[], true, &[], 0)
10162    }
10163
10164    /// **Bounded-export phase 2: eviction is byte-transparent, and progress is
10165    /// guaranteed at any cap.** A cap far below the decompressed dictionary
10166    /// forces the chunk cache to evict during a full dump, yet the dump is
10167    /// byte-for-byte the unlimited-cap dump — the `Arc` handout keeps every
10168    /// chunk alive across its one decode, so eviction never drops bytes in use.
10169    /// A 1-byte cap (smaller than any single chunk, so every insert evicts
10170    /// itself immediately) still completes with identical output: it is the
10171    /// handout invariant, not residency, that makes a window make progress.
10172    #[test]
10173    fn tiny_dict_cache_evicts_yet_dump_is_byte_identical() {
10174        use crate::reader::SliceReader;
10175        let leaked: &'static [u8] = Box::leak(build_multichunk_image().into_boxed_slice());
10176        let dump = |cap: u64| -> (
10177            Vec<(String, String, String)>,
10178            crate::chunk_cache::CacheStats,
10179        ) {
10180            let rete =
10181                Rete::open_ranged_lazy_with(SliceReader::new(leaked), HEADER_LEN as u64).unwrap();
10182            rete.dict.set_cache_cap(cap);
10183            let mut out = Vec::new();
10184            rete.dump_filtered_each(None, None, None, None, |s, p, o| {
10185                out.push((s.to_string(), p.to_string(), o.to_string()))
10186            });
10187            assert!(
10188                !rete.index_incomplete(),
10189                "no lazy fault should fail at cap {cap}"
10190            );
10191            let stats = rete.dict_cache_stats();
10192            (out, stats)
10193        };
10194
10195        let (unlimited, us) = dump(u64::MAX);
10196        assert_eq!(unlimited.len(), 4001, "every triple dumped");
10197        assert_eq!(us.evictions, 0, "unlimited cap evicts nothing");
10198
10199        // A cap far below the decompressed dictionary: eviction happens, output
10200        // unchanged.
10201        let (tiny, ts) = dump(128 * 1024);
10202        assert_eq!(tiny, unlimited, "128 KiB-cap dump differs from unlimited");
10203        assert!(ts.evictions > 0, "a tiny cap must evict (stats {ts:?})");
10204
10205        // Progress guarantee: a 1-byte cap still completes, byte-identical.
10206        let (starved, ss) = dump(1);
10207        assert_eq!(starved, unlimited, "1-byte-cap dump differs from unlimited");
10208        assert!(
10209            ss.evictions > 0,
10210            "1-byte cap evicts every insert (stats {ss:?})"
10211        );
10212    }
10213
10214    /// **The budget pays for the open's fixed directories.** A ranged open holds
10215    /// one descriptor per dictionary chunk and per index tile for its whole
10216    /// life, outside every cache — 88 MB on the 14.75 GiB openaire-relation,
10217    /// 224 MB on epfl-infoscience. `set_memory_budget` used to give the caches
10218    /// the whole budget on top of that; it now measures the directories
10219    /// ([`Rete::fixed_overhead`]) and takes them out of the dictionary share, so
10220    /// caps + directories stay within the total. This checks the arithmetic on
10221    /// a many-chunk file, the 64 MiB floor, the unlimited pass-through, and that
10222    /// a dump under the capped budget keeps the dictionary cache within its
10223    /// (net) share.
10224    #[test]
10225    fn memory_budget_charges_the_fixed_directories() {
10226        use crate::reader::SliceReader;
10227        let leaked: &'static [u8] = Box::leak(build_multichunk_image().into_boxed_slice());
10228        let rete =
10229            Rete::open_ranged_lazy_with(SliceReader::new(leaked), HEADER_LEN as u64).unwrap();
10230
10231        let fixed = rete.fixed_overhead();
10232        // Every chunk and every tile has a descriptor; a multi-chunk dictionary
10233        // and a 64-triple tile budget make both directories non-trivial.
10234        assert!(
10235            fixed.dict_directory > 0 && fixed.index_directory > 0,
10236            "{fixed:?}"
10237        );
10238        assert_eq!(fixed.named_graphs, 0, "no named graphs in this file");
10239        assert_eq!(
10240            fixed.total(),
10241            fixed.dict_directory + fixed.index_directory + fixed.named_graphs
10242        );
10243
10244        // A budget well above the floor: the dictionary share is the plain
10245        // split less the directories, exactly.
10246        let total: u64 = 512 << 20;
10247        let applied = rete.set_memory_budget(Some(total));
10248        let plain = split_memory_budget(total);
10249        assert_eq!(applied.dict_cache, plain.dict_cache - fixed.total());
10250        assert_eq!(rete.dict_cache_cap(), applied.dict_cache);
10251        assert_eq!(applied.tile_cache, plain.tile_cache);
10252        assert_eq!(applied.block_cache, plain.block_cache);
10253        assert!(
10254            applied.dict_cache + applied.tile_cache + applied.block_cache + fixed.total() <= total,
10255            "caps + fixed directories exceed the budget: {applied:?} + {fixed:?} > {total}"
10256        );
10257
10258        // A budget whose share would fall under the floor keeps the floor.
10259        let applied = rete.set_memory_budget(Some(1 << 20));
10260        assert_eq!(applied.dict_cache, MIN_DICT_CACHE_CAP);
10261
10262        // Unlimited stays unlimited: nothing is subtracted from u64::MAX.
10263        let applied = rete.set_memory_budget(None);
10264        assert_eq!(applied.dict_cache, u64::MAX);
10265        assert_eq!(rete.dict_cache_cap(), u64::MAX);
10266
10267        // Under a cap smaller than the decompressed dictionary, a full dump
10268        // keeps the cache within it.
10269        let rete =
10270            Rete::open_ranged_lazy_with(SliceReader::new(leaked), HEADER_LEN as u64).unwrap();
10271        rete.dict.set_cache_cap(128 * 1024);
10272        let mut rows = 0usize;
10273        rete.dump_filtered_each(None, None, None, None, |_, _, _| rows += 1);
10274        assert_eq!(rows, 4001);
10275        let s = rete.dict_cache_stats();
10276        assert!(s.evictions > 0, "the cap must bind: {s:?}");
10277        // One oversized chunk may sit alone above the cap (the handout
10278        // invariant); otherwise residency is within it.
10279        assert!(
10280            s.entries <= 1 || s.resident_bytes <= 128 * 1024,
10281            "dictionary cache over its cap after the dump: {s:?}"
10282        );
10283    }
10284
10285    /// **Bounded-export phase 2: per-graph index release.** After a named
10286    /// graph's slot is exported, [`Rete::release_named_graph`] drops its decoded
10287    /// index (and its tiles) so a many-graph dump holds one graph's index at a
10288    /// time, not all of them. The graph re-opens lazily if touched again.
10289    #[test]
10290    fn releasing_a_named_graph_drops_and_reopens_its_index() {
10291        use crate::index::PermSet;
10292        use crate::reader::SliceReader;
10293        let mut db = DictionaryBuilder::new();
10294        let all: Vec<(String, String, String)> = (0..40u32)
10295            .map(|i| {
10296                (
10297                    format!("<http://ex/s{}>", i % 7),
10298                    format!("<http://ex/p{}>", i % 3),
10299                    format!("<http://ex/o{}>", i % 5),
10300                )
10301            })
10302            .collect();
10303        for (s, p, o) in &all {
10304            db.observe(s, p, o);
10305        }
10306        let dict = db.build();
10307        let mk = |take: fn(u32) -> bool| {
10308            let mut g = GraphIndexBuilder::new()
10309                .with_tile_budget(16)
10310                .with_perms(PermSet::ALL);
10311            for (i, (s, p, o)) in all.iter().enumerate() {
10312                if take(i as u32) {
10313                    g.push(dict.encode(s, p, o).unwrap());
10314                }
10315            }
10316            g.build()
10317        };
10318        let def = mk(|_| true);
10319        let g1 = mk(|i| i % 2 == 0);
10320        let g2 = mk(|i| i % 3 == 0);
10321        let image = write_dataset(
10322            &dict,
10323            &def,
10324            &[
10325                ("<http://ex/g1>".to_string(), g1),
10326                ("<http://ex/g2>".to_string(), g2),
10327            ],
10328            true,
10329            &[],
10330            0,
10331        );
10332        let leaked: &'static [u8] = Box::leak(image.into_boxed_slice());
10333        let mut rete =
10334            Rete::open_ranged_lazy_with(SliceReader::new(leaked), HEADER_LEN as u64).unwrap();
10335
10336        let opened = |r: &Rete| -> usize {
10337            match &r.named_graphs {
10338                NamedGraphsSlot::Lazy(l) => {
10339                    let mut n = 0;
10340                    l.for_each_opened(|_| n += 1);
10341                    n
10342                }
10343                _ => 0,
10344            }
10345        };
10346
10347        assert_eq!(opened(&rete), 0, "nothing opened before first touch");
10348        // Touching g1 opens its index.
10349        assert!(rete.graph_index("<http://ex/g1>").is_some());
10350        assert_eq!(opened(&rete), 1, "g1 opened");
10351        // Release it: the index is dropped.
10352        rete.release_named_graph("<http://ex/g1>");
10353        assert_eq!(opened(&rete), 0, "g1 released");
10354        // Releasing an unknown / already-released graph is a no-op.
10355        rete.release_named_graph("<http://ex/absent>");
10356        rete.release_named_graph("<http://ex/g1>");
10357        assert_eq!(opened(&rete), 0);
10358        // The graph re-opens lazily on the next access.
10359        assert!(rete.graph_index("<http://ex/g1>").is_some());
10360        assert_eq!(opened(&rete), 1, "g1 re-opened after release");
10361    }
10362
10363    /// **The filtered-dump correctness invariant** (#117 point 1): for every
10364    /// graph and every one of the eight bound/unbound shapes,
10365    /// `dump_filtered_each` must return *exactly* the quads `dump_each` returns,
10366    /// filtered — no more, and above all no fewer.
10367    ///
10368    /// "No fewer" is the whole risk. The speedup comes from *not fetching* tiles
10369    /// a synopsis says cannot match, and a pruning bug does not crash, it
10370    /// silently returns a shorter answer. So the oracle here is the unfiltered
10371    /// dump with a plain Rust `retain` over it — the thing the filter is meant
10372    /// to be a cheaper spelling of — and it is checked on every open path, since
10373    /// the eager path prunes with a different code path than the lazy one.
10374    ///
10375    /// It runs on a `PermSet::CORE` (`--permutations 3`) file as well as a
10376    /// six-permutation one. Every shape routes inside CORE — `perm_routing_never_leaves_core`
10377    /// proves the routing table does — so the two must agree row for row, and if
10378    /// a future shape ever routed outside CORE this test fails on the three-perm
10379    /// build rather than silently returning nothing from an absent section.
10380    #[test]
10381    fn filtered_dump_is_exactly_the_unfiltered_dump_filtered() {
10382        use crate::index::PermSet;
10383        use crate::reader::SliceReader;
10384
10385        let terms = |i: u32| {
10386            (
10387                format!("<http://ex/s{}>", i % 11),
10388                format!("<http://ex/p{}>", i % 3),
10389                // Mix IRIs and literals in the object column: a literal lives in
10390                // a different dictionary section than a node, and the object
10391                // filter has to resolve through the right one.
10392                if i.is_multiple_of(4) {
10393                    format!("\"lit {}\"@en", i % 5)
10394                } else {
10395                    format!("<http://ex/o{}>", i % 7)
10396                },
10397            )
10398        };
10399        let build = |perms: PermSet| {
10400            let mut db = DictionaryBuilder::new();
10401            let all: Vec<_> = (0..90u32).map(terms).collect();
10402            for (s, p, o) in &all {
10403                db.observe(s, p, o);
10404            }
10405            let dict = db.build();
10406            let mut def = GraphIndexBuilder::new()
10407                .with_tile_budget(16)
10408                .with_perms(perms);
10409            let mut g1 = GraphIndexBuilder::new()
10410                .with_tile_budget(16)
10411                .with_perms(perms);
10412            for (i, (s, p, o)) in all.iter().enumerate() {
10413                let t = dict.encode(s, p, o).unwrap();
10414                // Deliberately overlapping, not disjoint: every second triple is
10415                // in both graphs, so a filter that leaked across graphs would
10416                // still look plausible on counts alone.
10417                if i % 2 == 0 {
10418                    g1.push(t);
10419                }
10420                def.push(t);
10421            }
10422            write_dataset(
10423                &dict,
10424                &def.build(),
10425                &[("<http://ex/g1>".to_string(), g1.build())],
10426                true,
10427                &[],
10428                0,
10429            )
10430        };
10431
10432        let shapes: Vec<(Option<&str>, Option<&str>, Option<&str>)> = vec![
10433            (None, None, None),
10434            (Some("<http://ex/s3>"), None, None),
10435            (None, Some("<http://ex/p1>"), None),
10436            (None, None, Some("<http://ex/o5>")),
10437            (None, None, Some("\"lit 0\"@en")),
10438            (Some("<http://ex/s3>"), Some("<http://ex/p0>"), None),
10439            (Some("<http://ex/s4>"), None, Some("<http://ex/o4>")),
10440            (None, Some("<http://ex/p1>"), Some("<http://ex/o5>")),
10441            (
10442                Some("<http://ex/s3>"),
10443                Some("<http://ex/p0>"),
10444                Some("<http://ex/o3>"),
10445            ),
10446            // Terms the dictionary has never seen, in each position: these must
10447            // return nothing, not everything (an unresolvable bound term must
10448            // not silently degrade to "unbound").
10449            (Some("<http://ex/nope>"), None, None),
10450            (None, Some("<http://ex/nope>"), None),
10451            (None, None, Some("<http://ex/nope>")),
10452        ];
10453
10454        let mut any_pruned = false;
10455        for perms in [PermSet::ALL, PermSet::CORE] {
10456            let image = build(perms);
10457            let leaked: &'static [u8] = Box::leak(image.clone().into_boxed_slice());
10458            let opens: Vec<(&str, Rete)> = vec![
10459                ("resident", Rete::open(&image).unwrap()),
10460                (
10461                    "ranged",
10462                    Rete::open_ranged(&SliceReader::new(leaked)).unwrap(),
10463                ),
10464                (
10465                    "lazy",
10466                    Rete::open_ranged_lazy_with(SliceReader::new(leaked), HEADER_LEN as u64)
10467                        .unwrap(),
10468                ),
10469            ];
10470            for graph in [None, Some("<http://ex/g1>"), Some("<http://ex/absent>")] {
10471                // The oracle: the unfiltered dump, filtered in Rust.
10472                let full: Vec<TermTriple> = {
10473                    let mut v = Vec::new();
10474                    Rete::open(&image).unwrap().dump_each(graph, |s, p, o| {
10475                        v.push((s.to_string(), p.to_string(), o.to_string()))
10476                    });
10477                    v.sort();
10478                    v
10479                };
10480                for &(s, p, o) in &shapes {
10481                    let mut want = full.clone();
10482                    want.retain(|(ts, tp, to)| {
10483                        s.is_none_or(|x| x == ts)
10484                            && p.is_none_or(|x| x == tp)
10485                            && o.is_none_or(|x| x == to)
10486                    });
10487                    for (path, rete) in &opens {
10488                        let mut got = Vec::new();
10489                        rete.dump_filtered_each(graph, s, p, o, |s, p, o| {
10490                            got.push((s.to_string(), p.to_string(), o.to_string()))
10491                        });
10492                        got.sort();
10493                        assert_eq!(
10494                            got,
10495                            want,
10496                            "{path}/{perms:?} graph={graph:?} pattern={:?} — filtered dump \
10497                             is not the unfiltered dump filtered",
10498                            (s, p, o)
10499                        );
10500                        assert!(!rete.index_incomplete(), "{path}: a fetch failed");
10501                    }
10502                    // …and the plan must agree with what the dump actually did:
10503                    // an empty answer means either no scan at all, or a scan
10504                    // whose admitted tiles simply held nothing.
10505                    let plan = opens[2].1.dump_plan(graph, s, p, o);
10506                    if graph == Some("<http://ex/absent>")
10507                        || [s, p, o].contains(&Some("<http://ex/nope>"))
10508                    {
10509                        assert!(
10510                            plan.scan.is_none(),
10511                            "an unmatchable dump still planned a scan: {:?}",
10512                            (graph, s, p, o)
10513                        );
10514                    } else {
10515                        let scan = plan.scan.expect("a matchable dump plans a scan");
10516                        assert!(
10517                            PermSet::CORE.contains(scan.permutation),
10518                            "a dump routed to {:?}, outside PermSet::CORE — a \
10519                             --permutations 3 file has no such section",
10520                            scan.permutation
10521                        );
10522                        assert!(scan.tiles_admitted <= scan.tiles_routed);
10523                        assert!(scan.tiles_routed <= scan.tiles_total);
10524                        assert!(scan.tile_bytes <= scan.section_bytes);
10525                        if scan.tiles_admitted < scan.tiles_total {
10526                            any_pruned = true;
10527                        }
10528                    }
10529                }
10530            }
10531        }
10532        assert!(
10533            any_pruned,
10534            "no shape pruned a single tile — the fixture is too small to be \
10535             measuring anything"
10536        );
10537    }
10538
10539    /// The point of the filter: a predicate-scoped dump of a lazily range-read
10540    /// file must **fetch less** than the same graph's full dump, not merely
10541    /// return less.
10542    ///
10543    /// The fixture is the shape the win shows on — many tiles, and a predicate
10544    /// whose triples live in a minority of them — so the assertion can be about
10545    /// bytes rather than about the code path being taken on faith. Measured on
10546    /// the real thing (`cordis.rete`, 801 MB): one predicate of one named graph
10547    /// read 16.0 MB where the graph's full dump read 375.8 MB.
10548    #[test]
10549    fn a_predicate_scoped_dump_fetches_less_than_the_graph() {
10550        use crate::reader::{CountingReader, SliceReader};
10551
10552        // 200 subjects × 20 predicates, tiled small so there are many tiles, and
10553        // one distinct object per cell so that a predicate slice needs a
10554        // twentieth of the object dictionary rather than all of it — otherwise
10555        // the dictionary, which this change does not prune, swamps the index
10556        // saving the test is about.
10557        let mut db = DictionaryBuilder::new();
10558        let mut all = Vec::new();
10559        for s in 0..200u32 {
10560            for p in 0..20u32 {
10561                let t = (
10562                    format!("<http://ex/s/{s:04}>"),
10563                    format!("<http://ex/p/{p:02}>"),
10564                    format!("\"object {s:04}/{p:02}\""),
10565                );
10566                db.observe(&t.0, &t.1, &t.2);
10567                all.push(t);
10568            }
10569        }
10570        let dict = db.build();
10571        let mut ib = GraphIndexBuilder::new().with_tile_budget(64);
10572        for (s, p, o) in &all {
10573            ib.push(dict.encode(s, p, o).unwrap());
10574        }
10575        let image = write_dataset(&dict, &ib.build(), &[], false, &[], 0);
10576        let leaked: &'static [u8] = Box::leak(image.clone().into_boxed_slice());
10577
10578        // Bytes faulted past the open, for one dump.
10579        let probe = |s: Option<&str>, p: Option<&str>, o: Option<&str>| -> (usize, u64) {
10580            let reader = std::sync::Arc::new(CountingReader::new(SliceReader::new(leaked)));
10581            let lazy = Rete::open_ranged_lazy_with(reader.clone(), HEADER_LEN as u64).unwrap();
10582            let before = reader.bytes_read();
10583            let mut n = 0usize;
10584            lazy.dump_filtered_each(None, s, p, o, |_, _, _| n += 1);
10585            assert!(!lazy.index_incomplete());
10586            (n, reader.bytes_read() - before)
10587        };
10588
10589        let (full_rows, full_bytes) = probe(None, None, None);
10590        let (pred_rows, pred_bytes) = probe(None, Some("<http://ex/p/07>"), None);
10591        let (subj_rows, subj_bytes) = probe(Some("<http://ex/s/0021>"), None, None);
10592        assert_eq!(full_rows, 4000);
10593        assert_eq!(pred_rows, 200, "one predicate covers every subject once");
10594        assert_eq!(subj_rows, 20);
10595        // True in any build: a slice fetches strictly less than the graph.
10596        assert!(
10597            pred_bytes < full_bytes && subj_bytes < full_bytes,
10598            "a slice fetched {pred_bytes} / {subj_bytes} B where the whole graph fetched \
10599             {full_bytes} B — the scan is not being pruned at all"
10600        );
10601        // The SIZE of the win is a property of the FILE, not of this change: it is
10602        // the index share of what a dump reads, and the dictionary — which no
10603        // filter prunes — is the rest. With the default `compression` feature the
10604        // dictionary chunk bodies are zstd-compressed and the slice costs about a
10605        // tenth of the graph (measured here: 1,565 B for the predicate and 864 B
10606        // for the subject, against 15,714 B). Built `--no-default-features` those
10607        // same chunks are stored RAW, so resolving 200 rows' terms dominates and
10608        // the identical, identically-pruned scan costs 22,225 B of 40,165 — 55%.
10609        // Asserting the strong ratio unconditionally would be asserting that zstd
10610        // is enabled; the index-side prune is checked below and holds either way.
10611        #[cfg(feature = "compression")]
10612        assert!(
10613            pred_bytes * 4 < full_bytes && subj_bytes * 4 < full_bytes,
10614            "a 1-in-20 predicate slice fetched {pred_bytes} B and one subject \
10615             {subj_bytes} B where the whole graph fetched {full_bytes} B"
10616        );
10617
10618        // And the plan predicted the index side of it without fetching a tile.
10619        let reader = std::sync::Arc::new(CountingReader::new(SliceReader::new(leaked)));
10620        let lazy = Rete::open_ranged_lazy_with(reader.clone(), HEADER_LEN as u64).unwrap();
10621        let before = reader.bytes_read();
10622        let plan = lazy
10623            .dump_plan(None, None, Some("<http://ex/p/07>"), None)
10624            .scan
10625            .expect("a known predicate plans a scan");
10626        assert!(
10627            reader.bytes_read() - before < 64 * 1024,
10628            "planning fetched {} B — it is supposed to read directories, not tiles",
10629            reader.bytes_read() - before
10630        );
10631        assert!(plan.tiles_admitted < plan.tiles_total);
10632        // The index side on its own, with the dictionary's share taken out: one
10633        // predicate of twenty routes to a handful of POS tiles. (Measured here:
10634        // the whole dump faults 15,714 B, the predicate slice 1,565 B and one
10635        // subject 864 B — a 10x and an 18x, most of what is left being the
10636        // dictionary this change does not prune.)
10637        assert!(
10638            plan.tile_bytes * 5 < plan.section_bytes,
10639            "one predicate of twenty planned {} B of a {} B section",
10640            plan.tile_bytes,
10641            plan.section_bytes
10642        );
10643    }
10644
10645    #[test]
10646    fn query_quads_tags_every_graph() {
10647        let mut db = DictionaryBuilder::new();
10648        for (s, p, o) in [("Alice", "knows", "Bob"), ("Alice", "knows", "Dave")] {
10649            db.observe(s, p, o);
10650        }
10651        let dict = db.build();
10652        let mut def = GraphIndexBuilder::new();
10653        def.push(dict.encode("Alice", "knows", "Bob").unwrap());
10654        let mut g1 = GraphIndexBuilder::new();
10655        g1.push(dict.encode("Alice", "knows", "Dave").unwrap());
10656        let named = vec![("http://ex/g1".to_string(), g1.build())];
10657        let bytes = write_dataset(&dict, &def.build(), &named, true, &[], 0);
10658        let rete = Rete::open(&bytes).unwrap();
10659
10660        // `Alice knows ?` spans both graphs; each match carries its graph tag.
10661        let quads = rete.query_quads(Some("Alice"), Some("knows"), None);
10662        assert_eq!(quads.len(), 2);
10663        assert_eq!(
10664            quads[0],
10665            (("Alice".into(), "knows".into(), "Bob".into()), None)
10666        );
10667        assert_eq!(
10668            quads[1],
10669            (
10670                ("Alice".into(), "knows".into(), "Dave".into()),
10671                Some("http://ex/g1".to_string())
10672            )
10673        );
10674
10675        // A bound term absent from the dictionary yields nothing, in any graph.
10676        assert!(rete.query_quads(Some("Nobody"), None, None).is_empty());
10677    }
10678
10679    #[test]
10680    fn pyramid_meta_round_trips_in_file() {
10681        let rete = Rete::open(&build_image()).unwrap();
10682        let pyr = rete.pyramid().expect("file has a pyramid");
10683        // Summary covers all 3 triples by count; tiles are not stored in v0.
10684        let total: u32 = pyr.summary.iter().map(|e| e.count).sum();
10685        assert_eq!(total, 3);
10686        assert!(!pyr.summary.is_empty());
10687        assert!(pyr.tiles.is_empty());
10688    }
10689
10690    #[test]
10691    fn schema_pyramid_round_trips_through_file_index_free() {
10692        use crate::reader::{CountingReader, SliceReader};
10693        let sub = "<http://www.w3.org/2000/01/rdf-schema#subClassOf>";
10694        let q = |s: &str, p: &str, o: &str| {
10695            (s.to_string(), p.to_string(), o.to_string(), None::<String>)
10696        };
10697        // Astronomer ⊑ Scientist ⊑ Person ⊑ Agent, instances at the leaves.
10698        let quads = vec![
10699            q("<a>", RDF_TYPE, "<Astronomer>"),
10700            q("<b>", RDF_TYPE, "<Astronomer>"),
10701            q("<c>", RDF_TYPE, "<Person>"),
10702            q("<Astronomer>", sub, "<Scientist>"),
10703            q("<Scientist>", sub, "<Person>"),
10704            q("<Person>", sub, "<Agent>"),
10705            q("<a>", "<knows>", "<b>"),
10706            q("<b>", "<knows>", "<c>"),
10707        ];
10708        let (bytes, _) =
10709            crate::ingest::assemble_dataset_with_opts(quads, true, false, None, |_, _| Vec::new());
10710
10711        // The v2 schema pyramid round-trips through the built file.
10712        let rete = Rete::open(&bytes).unwrap();
10713        let pyr = rete.pyramid().expect("pyramid present");
10714        assert!(!pyr.level_rollups.is_empty(), "schema pyramid shipped");
10715        assert!(pyr
10716            .class_hierarchy
10717            .iter()
10718            .any(|n| n.class == "<Agent>" && n.depth == 0));
10719
10720        // It reads index-free: a SummaryView open never touches the index section.
10721        let r = CountingReader::new(SliceReader::new(&bytes));
10722        let view = SummaryView::open_ranged(&r).unwrap().unwrap();
10723        assert!(view.level_count() >= 2, "multi-level pyramid");
10724        let coarse = view.level_rollup(0).unwrap();
10725        assert!(
10726            coarse.classes.iter().any(|(c, _)| c == "<Agent>"),
10727            "coarsest level rolls up to the root Agent"
10728        );
10729        let h = Header::from_bytes(&bytes).unwrap();
10730        assert!(
10731            r.bytes_read() <= bytes.len() as u64 - h.root_dir_len,
10732            "summary read {} bytes; the {}-byte index section must be skipped",
10733            r.bytes_read(),
10734            h.root_dir_len
10735        );
10736    }
10737
10738    #[test]
10739    fn predicate_totals_from_summary_only() {
10740        use crate::reader::SliceReader;
10741        // build_image: 2 `knows` triples + 1 `age` triple.
10742        let bytes = build_image();
10743        let reader = SliceReader::new(&bytes);
10744        let view = SummaryView::open_ranged(&reader).unwrap().unwrap();
10745        assert_eq!(view.predicate_total("knows"), 2);
10746        assert_eq!(view.predicate_total("age"), 1);
10747        assert_eq!(view.predicate_total("missing"), 0);
10748        let totals = view.predicate_totals();
10749        assert_eq!(totals[0], ("knows".to_string(), 2)); // sorted by count desc
10750    }
10751
10752    #[test]
10753    fn query_patterns_resolve_to_terms() {
10754        let rete = Rete::open(&build_image()).unwrap();
10755
10756        // All triples.
10757        assert_eq!(rete.query(None, None, None).len(), 3);
10758
10759        // Subject bound.
10760        let mut alice = rete.query(Some("Alice"), None, None);
10761        alice.sort();
10762        assert_eq!(
10763            alice,
10764            vec![
10765                ("Alice".into(), "age".into(), "30".into()),
10766                ("Alice".into(), "knows".into(), "Bob".into()),
10767            ]
10768        );
10769
10770        // Predicate bound.
10771        assert_eq!(rete.query(None, Some("knows"), None).len(), 2);
10772
10773        // Full triple, present and absent.
10774        assert_eq!(
10775            rete.query(Some("Bob"), Some("knows"), Some("Carol")),
10776            vec![("Bob".into(), "knows".into(), "Carol".into())]
10777        );
10778        assert!(rete.query(Some("Nobody"), None, None).is_empty());
10779        assert!(rete.query(None, Some("likes"), None).is_empty());
10780    }
10781
10782    #[test]
10783    fn query_provenance_reports_terms_ids_sections_and_index_choice() {
10784        let bytes = build_image();
10785        let rete = Rete::open(&bytes).unwrap();
10786
10787        let mut matches = rete.query_with_provenance(None, Some("knows"), None);
10788        matches.sort_by(|a, b| a.terms.cmp(&b.terms));
10789
10790        assert_eq!(matches.len(), 2);
10791        assert_eq!(
10792            matches[0].terms,
10793            ("Alice".into(), "knows".into(), "Bob".into())
10794        );
10795        assert_eq!(
10796            matches[0].ids,
10797            rete.dictionary().encode("Alice", "knows", "Bob").unwrap()
10798        );
10799        assert_eq!(matches[0].graph.as_deref(), None);
10800        assert_eq!(
10801            matches[0].matched_pattern,
10802            (None, Some(matches[0].ids.1), None)
10803        );
10804        assert_eq!(
10805            matches[0].index_permutation,
10806            crate::index::IndexPermutation::Pos
10807        );
10808
10809        let h = rete.header();
10810        assert_eq!(matches[0].dictionary_range.offset, h.dictionary_offset);
10811        assert_eq!(matches[0].dictionary_range.len, h.dictionary_len);
10812        assert_eq!(matches[0].index_range.offset, h.root_dir_offset);
10813        assert_eq!(matches[0].index_range.len, h.root_dir_len);
10814        assert!(
10815            matches[0].index_section_range.offset > h.root_dir_offset,
10816            "POS is section 1, so its payload starts after the container header and SPO payload"
10817        );
10818        assert!(matches[0].index_section_range.len > 0);
10819        assert!(matches[0].index_section_range.end() <= matches[0].index_range.end());
10820        assert!(matches[0].index_section_range.len < matches[0].index_range.len);
10821        assert_eq!(
10822            matches[0].pyramid_range.as_ref().map(|r| (r.offset, r.len)),
10823            Some((h.pyramid_meta_offset, h.pyramid_meta_len))
10824        );
10825        // Tiled (v0.2) files report the physical tile holding the match; its
10826        // compressed byte range nests inside the selected section payload.
10827        let tile_range = matches[0].tile_range.expect("tiled file reports a tile");
10828        assert!(matches[0]
10829            .tile
10830            .as_deref()
10831            .unwrap()
10832            .starts_with(matches[0].index_permutation.name()));
10833        assert!(matches[0].index_section_range.offset <= tile_range.offset);
10834        assert!(tile_range.end() <= matches[0].index_section_range.end());
10835    }
10836
10837    /// Build an in-memory `.rete` with `n` labeled subjects: each carries an
10838    /// `rdfs:label` literal drawn from `WORDS` (a word prefix selects ~1/|WORDS|
10839    /// of them) plus one extra edge so the subject has a degree to rank by.
10840    fn build_labeled(n: usize) -> Vec<u8> {
10841        const LABEL: &str = "<http://www.w3.org/2000/01/rdf-schema#label>";
10842        const WORDS: &[&str] = &[
10843            "alanine",
10844            "benzene",
10845            "glucose",
10846            "dextrose",
10847            "ethanol",
10848            "formate",
10849            "heptane",
10850            "isoleucine",
10851        ];
10852        let triples: Vec<(String, String, String)> = (0..n)
10853            .flat_map(|i| {
10854                let s = format!("<http://ex/e{i}>");
10855                let w = WORDS[i % WORDS.len()];
10856                [
10857                    (s.clone(), LABEL.to_string(), format!("\"{w}-{i:06}\"")),
10858                    (
10859                        s,
10860                        "<http://ex/p>".to_string(),
10861                        format!("<http://ex/c{}>", i % 64),
10862                    ),
10863                ]
10864            })
10865            .collect();
10866        let mut db = DictionaryBuilder::new();
10867        for (s, p, o) in &triples {
10868            db.observe(s, p, o);
10869        }
10870        let dict = db.build();
10871        let ids: Vec<(u32, u32, u32)> = triples
10872            .iter()
10873            .map(|(s, p, o)| dict.encode(s, p, o).unwrap())
10874            .collect();
10875        let mut ib = GraphIndexBuilder::new();
10876        for &t in &ids {
10877            ib.push(t);
10878        }
10879        let (meta, levels) = build_pyramid_meta(&dict, &ids, DEFAULT_TILE_BUDGET);
10880        write_file(&dict, &ib.build(), false, &meta, levels)
10881    }
10882
10883    #[test]
10884    fn prefix_search_matches_a_filter_scan() {
10885        // 800 < the 8192 label-index cap, so the index is COMPLETE — every label
10886        // is present and the two paths must return the exact same subject set.
10887        let bytes = build_labeled(800);
10888        let rete = Rete::open(&bytes).unwrap();
10889        let idx_subjects: std::collections::BTreeSet<String> = rete
10890            .prefix_search("glucose", 10_000)
10891            .into_iter()
10892            .map(|(_label, subject)| subject)
10893            .collect();
10894        // The same selection via a SPARQL FILTER scan over every label literal.
10895        let q = "SELECT ?s WHERE { ?s <http://www.w3.org/2000/01/rdf-schema#label> ?l \
10896                 FILTER(STRSTARTS(LCASE(?l), \"glucose\")) }";
10897        let crate::QueryOutput::Select(_, rows) = crate::eval_query(&rete, q).unwrap() else {
10898            panic!("expected SELECT");
10899        };
10900        let scan_subjects: std::collections::BTreeSet<String> =
10901            rows.iter().map(|r| r.get("s").cloned().unwrap()).collect();
10902        assert_eq!(idx_subjects, scan_subjects, "index agrees with the scan");
10903        assert_eq!(
10904            idx_subjects.len(),
10905            100,
10906            "800/8 words = 100 glucose-* labels"
10907        );
10908    }
10909
10910    /// A legacy (pyramid-block) label search resolves its hits' subjects in one
10911    /// batched dictionary read on a lazy open — the label-search twin of the
10912    /// full-text hits' batch — instead of one dependent read per chunk.
10913    #[test]
10914    fn legacy_label_hits_resolve_in_one_batched_read() {
10915        use std::sync::{Arc, Mutex};
10916        type CallLog = Arc<Mutex<Vec<Vec<(u64, u64)>>>>;
10917        #[derive(Clone)]
10918        struct Calls(Arc<Vec<u8>>, CallLog);
10919        impl crate::RangeReader for Calls {
10920            fn len(&self) -> u64 {
10921                self.0.len() as u64
10922            }
10923            fn read_at(&self, offset: u64, len: u64) -> std::io::Result<Vec<u8>> {
10924                self.1.lock().unwrap().push(vec![(offset, len)]);
10925                crate::SliceReader::new(&self.0).read_at(offset, len)
10926            }
10927            fn read_many(&self, ranges: &[(u64, u64)]) -> std::io::Result<Vec<Vec<u8>>> {
10928                self.1.lock().unwrap().push(ranges.to_vec());
10929                let r = crate::SliceReader::new(&self.0);
10930                ranges.iter().map(|&(o, l)| r.read_at(o, l)).collect()
10931            }
10932        }
10933        let bytes = build_labeled(20_000);
10934        let h = Header::from_bytes(&bytes).unwrap();
10935        let (d0, dl) = (h.dictionary_offset, h.dictionary_len);
10936        let want = Rete::open(&bytes).unwrap().prefix_search("glucose", 1000);
10937        assert_eq!(want.len(), 1000);
10938        for search_view in [false, true] {
10939            let calls = Calls(Arc::new(bytes.clone()), Arc::new(Mutex::new(Vec::new())));
10940            let got = if search_view {
10941                let view = SearchView::open_ranged_with(calls.clone(), HEADER_LEN as u64).unwrap();
10942                calls.1.lock().unwrap().clear();
10943                view.prefix_search("glucose", 1000)
10944            } else {
10945                let rete = Rete::open_ranged_lazy_with(calls.clone(), HEADER_LEN as u64).unwrap();
10946                calls.1.lock().unwrap().clear();
10947                rete.prefix_search("glucose", 1000)
10948            };
10949            assert_eq!(got, want, "search view = {search_view}");
10950            let dict_reads = calls
10951                .1
10952                .lock()
10953                .unwrap()
10954                .iter()
10955                .filter(|c| c.iter().any(|&(o, l)| o < d0 + dl && d0 < o + l))
10956                .count();
10957            assert_eq!(dict_reads, 1, "search view = {search_view}");
10958        }
10959    }
10960
10961    /// Latency: the binary-search label index vs the FILTER scan it replaces.
10962    /// Ignored by default (timing-sensitive); run with
10963    /// `cargo test -p rete-core -- --ignored --nocapture bench_prefix_search`.
10964    #[test]
10965    #[ignore]
10966    fn bench_prefix_search_vs_filter_scan() {
10967        use std::time::Instant;
10968        let n = 6000; // < the 8192 cap, so both paths return identical sets
10969        let bytes = build_labeled(n);
10970        let rete = Rete::open(&bytes).unwrap();
10971        let q = "SELECT ?s WHERE { ?s <http://www.w3.org/2000/01/rdf-schema#label> ?l \
10972                 FILTER(STRSTARTS(LCASE(?l), \"glucose\")) }";
10973        let reps = 200;
10974        let idx_n = rete.prefix_search("glucose", 100_000).len();
10975        let t = Instant::now();
10976        for _ in 0..reps {
10977            std::hint::black_box(rete.prefix_search("glucose", 100_000));
10978        }
10979        let idx_ms = t.elapsed().as_secs_f64() * 1000.0 / reps as f64;
10980        let t = Instant::now();
10981        for _ in 0..reps {
10982            let _ = std::hint::black_box(crate::eval_query(&rete, q).unwrap());
10983        }
10984        let scan_ms = t.elapsed().as_secs_f64() * 1000.0 / reps as f64;
10985        println!(
10986            "label prefix search over {n} labeled subjects ({idx_n} matches): \
10987             index {idx_ms:.4} ms vs FILTER scan {scan_ms:.3} ms ({:.0}× faster)",
10988            scan_ms / idx_ms
10989        );
10990    }
10991
10992    /// Latency: the TEXT_INDEX word search vs the `FILTER(CONTAINS(?l, …))` scan
10993    /// it replaces. Ignored by default (timing-sensitive); run with
10994    /// `cargo test -p rete-core -- --ignored --nocapture bench_text_search`.
10995    #[test]
10996    #[ignore]
10997    fn bench_text_search_vs_contains_scan() {
10998        use std::time::Instant;
10999        const LABEL: &str = "<http://www.w3.org/2000/01/rdf-schema#label>";
11000        const WORDS: &[&str] = &[
11001            "alanine",
11002            "benzene",
11003            "glucose",
11004            "dextrose",
11005            "ethanol",
11006            "formate",
11007            "heptane",
11008            "isoleucine",
11009        ];
11010        let n = 6000;
11011        let triples: Vec<(String, String, String)> = (0..n)
11012            .map(|i| {
11013                (
11014                    format!("<http://ex/e{i}>"),
11015                    LABEL.to_string(),
11016                    format!("\"{} sample number {i:06}\"", WORDS[i % WORDS.len()]),
11017                )
11018            })
11019            .collect();
11020        let mut db = DictionaryBuilder::new();
11021        for (s, p, o) in &triples {
11022            db.observe(s, p, o);
11023        }
11024        let dict = db.build();
11025        let ids: Vec<(u32, u32, u32)> = triples
11026            .iter()
11027            .map(|(s, p, o)| dict.encode(s, p, o).unwrap())
11028            .collect();
11029        let mut ib = GraphIndexBuilder::new();
11030        for &t in &ids {
11031            ib.push(t);
11032        }
11033        let ti = compute_text_index(&dict, &ids);
11034        let bytes = write_dataset_with_metadata(&dict, &ib.build(), &[], false, &[], 0, &[], &ti);
11035        let rete = Rete::open(&bytes).unwrap();
11036
11037        let q = "SELECT ?s WHERE { ?s <http://www.w3.org/2000/01/rdf-schema#label> ?l \
11038                 FILTER(CONTAINS(LCASE(?l), \"glucose\")) }";
11039        let reps = 200;
11040        let idx_n = rete.text_search(&["glucose"], None, 100_000).len();
11041        let t = Instant::now();
11042        for _ in 0..reps {
11043            std::hint::black_box(rete.text_search(&["glucose"], None, 100_000));
11044        }
11045        let idx_ms = t.elapsed().as_secs_f64() * 1000.0 / reps as f64;
11046        let t = Instant::now();
11047        for _ in 0..reps {
11048            let _ = std::hint::black_box(crate::eval_query(&rete, q).unwrap());
11049        }
11050        let scan_ms = t.elapsed().as_secs_f64() * 1000.0 / reps as f64;
11051        println!(
11052            "text search over {n} literals ({idx_n} matches): \
11053             index {idx_ms:.4} ms vs FILTER(CONTAINS) scan {scan_ms:.3} ms ({:.0}× faster)",
11054            scan_ms / idx_ms
11055        );
11056    }
11057
11058    /// Operational debugging harness for a REAL on-disk file (ignored; driven
11059    /// by env vars): step-by-step dump of a bound (p, o) POS routing.
11060    ///   RETE_DEBUG_FILE=<path.rete> RETE_DEBUG_P=<iri> RETE_DEBUG_O=<iri>
11061    #[test]
11062    #[ignore = "operational tool, driven by RETE_DEBUG_* env vars"]
11063    fn debug_bound_po_routing() {
11064        struct FR(std::fs::File);
11065        impl crate::RangeReader for FR {
11066            fn len(&self) -> u64 {
11067                self.0.metadata().map(|m| m.len()).unwrap_or(0)
11068            }
11069            fn read_at(&self, offset: u64, len: u64) -> std::io::Result<Vec<u8>> {
11070                use std::os::unix::fs::FileExt;
11071                let mut buf = vec![0u8; len as usize];
11072                self.0.read_exact_at(&mut buf, offset)?;
11073                Ok(buf)
11074            }
11075        }
11076        let path = std::env::var("RETE_DEBUG_FILE").expect("RETE_DEBUG_FILE");
11077        let p_iri = std::env::var("RETE_DEBUG_P").expect("RETE_DEBUG_P");
11078        let o_iri = std::env::var("RETE_DEBUG_O").expect("RETE_DEBUG_O");
11079        let rete = Rete::open_ranged_lazy_with(
11080            std::sync::Arc::new(FR(std::fs::File::open(&path).unwrap())),
11081            HEADER_LEN as u64,
11082        )
11083        .unwrap();
11084        let pid = rete.dict.predicate_id(&p_iri).expect("p resolves");
11085        let oid = rete.dict.object_id(&o_iri).expect("o resolves");
11086        eprintln!("pid={pid} oid={oid}");
11087        let pattern = (None, Some(pid), Some(oid));
11088        let perm = GraphIndex::best_permutation(pattern);
11089        eprintln!("best_permutation = {}", perm.name());
11090        let si = perm.section_index();
11091        let tiles = &rete.index.sections[si];
11092        eprintln!("section {} tiles = {}", perm.name(), tiles.len());
11093        let [pa, pb, pc] = perm.order_pattern(pattern);
11094        eprintln!("permuted pattern pa={pa:?} pb={pb:?} pc={pc:?}");
11095        let (start, end) = rete.index.tile_span(si, pa);
11096        eprintln!("tile_span = [{start}, {end}) -> {} tiles", end - start);
11097        let mut admitted = 0usize;
11098        for (ti, t) in tiles.iter().enumerate().take(end).skip(start) {
11099            if t.syn_admits(pb, pc) {
11100                admitted += 1;
11101                if admitted <= 10 {
11102                    let (lo, hi) = t.leading_range();
11103                    eprintln!("  admit tile {ti}: a=[{lo},{hi}] syn={:?}", t.syn);
11104                }
11105            }
11106        }
11107        eprintln!("admitted {admitted} tile(s) by synopsis");
11108        let n = rete.index.scan_iter(pattern).count();
11109        eprintln!("scan_iter matches = {n}");
11110        let hi_res = rete.query(None, Some(&p_iri), Some(&o_iri));
11111        eprintln!("high-level query matches = {}", hi_res.len());
11112    }
11113
11114    /// The header segment's LABEL and the span it describes must agree. They
11115    /// did not from v0.3 (which grew the header to 1 KiB) until this test: the
11116    /// label still said "fixed 128 bytes", the pre-v0.3 size, while the segment
11117    /// covered `HEADER_LEN` = 1024 — and that label is exactly what the file
11118    /// explorer shows someone inspecting a real file. Reading the size back out
11119    /// of the string is the point: it fails if anyone retypes the number
11120    /// instead of deriving it from `HEADER_LEN`.
11121    #[test]
11122    fn layout_header_label_matches_its_span() {
11123        let mut db = DictionaryBuilder::new();
11124        let mut triples = Vec::new();
11125        for i in 0..24u32 {
11126            let (s, p, o) = (
11127                format!("<http://ex/s{}>", i % 5),
11128                format!("<http://ex/p{}>", i % 3),
11129                format!("<http://ex/o{i}>"),
11130            );
11131            db.observe(&s, &p, &o);
11132            triples.push((s, p, o));
11133        }
11134        let dict = db.build();
11135        let ids: Vec<(u32, u32, u32)> = triples
11136            .iter()
11137            .map(|(s, p, o)| dict.encode(s, p, o).unwrap())
11138            .collect();
11139        let index = GraphIndexBuilder::from_triples(ids).build();
11140        let image = write_dataset_with_metadata(
11141            &dict,
11142            &index,
11143            &[],
11144            false,
11145            &[],
11146            0,
11147            br#"{"name":"layout-label-fixture"}"#,
11148            &[],
11149        );
11150        let rete = Rete::open(&image).unwrap();
11151
11152        let layout = rete.file_layout();
11153        let header = layout
11154            .iter()
11155            .find(|s| s.kind == "header")
11156            .expect("every layout starts with a header segment");
11157
11158        assert_eq!(
11159            header.offset, 0,
11160            "the header is the first thing in the file"
11161        );
11162        assert_eq!(
11163            header.len, HEADER_LEN as u64,
11164            "the header segment must span exactly HEADER_LEN"
11165        );
11166
11167        // Pull the number back out of the human-readable label and hold it to
11168        // the span. A hard-coded literal that drifts from HEADER_LEN dies here.
11169        let digits: String = header
11170            .label
11171            .chars()
11172            .skip_while(|c| !c.is_ascii_digit())
11173            .take_while(|c| c.is_ascii_digit())
11174            .collect();
11175        assert!(
11176            !digits.is_empty(),
11177            "header label {:?} states no size at all",
11178            header.label
11179        );
11180        assert_eq!(
11181            digits.parse::<u64>().unwrap(),
11182            header.len,
11183            "header label {:?} disagrees with the {}-byte span it describes",
11184            header.label,
11185            header.len
11186        );
11187
11188        // The layout must also stay sorted and start at byte zero, which is
11189        // what makes "the first segment is the header" meaningful.
11190        assert!(
11191            layout.windows(2).all(|w| w[0].offset <= w[1].offset),
11192            "file_layout must return segments sorted by offset"
11193        );
11194    }
11195}