trusty_memory/

kg_extract.rs

//! Deterministic KG triple extraction from drawer content.
//!
//! Why: Issue #97 — `memory_remember` should populate the knowledge graph
//! automatically so palaces with drawers always have a non-empty KG. Calling an
//! LLM on every write would blow up latency and require network access; a
//! deterministic heuristic stays fast and offline while still producing useful
//! triples for tag membership, key-phrase mentions, and obvious is-a / has-a /
//! works-at patterns. The visual graph view (the other half of #97) renders
//! whatever shows up here, so this pass is the data source for "every palace
//! has a graph".
//! What: A pure function `extract_triples` that takes drawer content + tags +
//! drawer id and returns a `Vec<Triple>` with `provenance = "auto:remember"`.
//! The current heuristics are tag→drawer, room→drawer, hashtag→drawer, and a
//! short pattern table (`X is a Y`, `X works at Y`, `X uses Y`, `X depends on
//! Y`). Drawer ids are encoded as `drawer:<uuid>` so the subject keeps a
//! stable, palace-unique identity that the graph view can dereference back
//! to the source drawer.
//! Test: `extract_triples_emits_tag_triples`,
//! `extract_triples_emits_hashtag_mentions`,
//! `extract_triples_extracts_is_a_pattern`,
//! `extract_triples_never_panics_on_empty_input`.

use crate::wordnet_pos::{self, WordNetPos};
use chrono::Utc;
use std::collections::HashSet;
use trusty_common::memory_core::store::kg::Triple;
use uuid::Uuid;

/// Default tags that cause a drawer to be skipped during auto-extraction.
///
/// Why: Drawers tagged with these labels are by definition non-factual project
/// knowledge (test fixtures, QA scaffolding, synthetic content) and should not
/// pollute the KG with noise triples.
/// What: A static slice of lowercase tag strings; matched case-insensitively
/// during extraction.
/// Test: `extract_triples_skips_denied_tags`.
pub const DEFAULT_DENY_TAGS: &[&str] = &["cross-project-qa", "test", "fixture"];

/// Configuration for a single extraction pass.
///
/// Why: Bundles per-run configuration so `extract_triples` can be called with
/// different deny-lists (e.g. the default prod list vs. an empty list in
/// integration tests) without changing the function signature. #5399 added the
/// POS table here rather than to a process-wide `OnceLock`, which CLAUDE.md
/// forbids: the table is a 16-byte `Copy` handle over `&'static str` with no
/// build step, so passing it costs nothing and every caller can substitute a
/// different one.
/// What: Contains a `deny_tags` slice and a [`WordNetPos`] lookup. The
/// extractor skips any drawer whose tags intersect the deny set, and consults
/// the lookup when deciding which token in a noun phrase is its head.
/// `#[non_exhaustive]` so the next field is not another breaking change.
/// Test: `extract_triples_skips_denied_tags`, `extract_triples_empty_deny_list`,
/// `a_caller_supplied_pos_table_is_the_one_consulted`.
#[derive(Debug, Clone)]
#[non_exhaustive]
pub struct KgExtractConfig<'a> {
    /// Tags that cause extraction to be skipped entirely. Compared
    /// case-insensitively against the drawer's tag list.
    pub deny_tags: &'a [&'a str],
    /// Part-of-speech membership used to find the head of a noun phrase.
    pub pos: WordNetPos,
}

impl Default for KgExtractConfig<'_> {
    fn default() -> Self {
        Self {
            deny_tags: DEFAULT_DENY_TAGS,
            pos: WordNetPos::shipped(),
        }
    }
}

/// Provenance tag stamped on every auto-extracted triple.
///
/// Why: Operators need a stable string to filter / retract the auto-extracted
/// subset without scanning content. Centralising the constant keeps every
/// emitter and the back-fill CLI in sync.
/// What: A `&'static str` containing the literal `auto:remember`.
/// Test: `extract_triples_stamps_provenance`.
pub const AUTO_PROVENANCE: &str = "auto:remember";

/// Confidence applied to auto-extracted triples.
///
/// Why: Heuristic extraction is not authoritative; downstream rankers can use
/// the confidence to prefer explicit `kg_assert` triples over auto-extracted
/// noise.
/// What: `0.6` — high enough to surface in queries, low enough to be
/// over-ridden by a manual `kg_assert` of the same `(subject, predicate)`.
/// Test: `extract_triples_uses_reduced_confidence`.
pub const AUTO_CONFIDENCE: f32 = 0.6;

/// Subject prefix used for drawer-identity triples.
///
/// Why: A stable, palace-unique identifier lets the graph view dereference a
/// node back to the source drawer (and the back-fill CLI dedupe by drawer).
/// What: `drawer:` — concatenated with the drawer UUID hyphenated form.
/// Test: every test in this module asserts the prefix.
pub const DRAWER_SUBJECT_PREFIX: &str = "drawer:";

/// Subject prefix used for tag entities.
///
/// Why: The KG enforces at most one active triple per `(subject, predicate)`,
/// so we can't emit `drawer:X has-tag t1; drawer:X has-tag t2` — the second
/// assert would close the first. By promoting each tag to its own subject
/// (`tag:t1`, `tag:t2`) we keep multiple tags as distinct edges and the graph
/// view gets natural tag-clusters around each drawer.
/// What: `tag:` — concatenated with the lower-cased tag string.
/// Test: `extract_triples_emits_tag_triples`.
pub const TAG_SUBJECT_PREFIX: &str = "tag:";

/// Subject prefix used for free-text mention entities.
///
/// Why: Same temporal-invariant reasoning as `TAG_SUBJECT_PREFIX`. Hashtag
/// mentions and other discovered topical terms become their own subjects so
/// multiple mentions per drawer survive the assert pipeline.
/// What: `topic:` — concatenated with the lower-cased term.
/// Test: `extract_triples_emits_hashtag_mentions`.
pub const TOPIC_SUBJECT_PREFIX: &str = "topic:";

/// Subject prefix used for room entities.
///
/// Why: A drawer can only sit in one room, but encoding the room as its own
/// subject keeps the graph topology consistent (all "discovered metadata"
/// entities live under prefixed namespaces) and lets multiple drawers from
/// the same room cluster around a shared room node.
/// What: `room:` — concatenated with the room label.
/// Test: `extract_triples_emits_tag_triples`.
pub const ROOM_SUBJECT_PREFIX: &str = "room:";

/// Build the drawer subject string used as the (s) for every per-drawer
/// triple emitted by this module.
///
/// Why: Centralises the `drawer:<uuid>` encoding so call sites cannot drift.
/// What: Returns `format!("{DRAWER_SUBJECT_PREFIX}{id}")`.
/// Test: covered by every extractor test.
pub fn drawer_subject(id: Uuid) -> String {
    format!("{DRAWER_SUBJECT_PREFIX}{id}")
}

/// Inputs to a single extraction pass.
///
/// Why: Bundling the inputs keeps `extract_triples` signature small and lets
/// us add new fields (e.g. drawer_type) without breaking call sites.
/// What: Plain data struct; all fields are borrowed so the caller keeps
/// ownership.
/// Test: indirectly via every test that constructs one.
#[derive(Debug, Clone)]
pub struct ExtractInput<'a> {
    pub drawer_id: Uuid,
    pub content: &'a str,
    pub tags: &'a [String],
    pub room: Option<&'a str>,
}

/// Run the deterministic heuristic extractor with default config.
///
/// Why: Convenience wrapper that uses [`KgExtractConfig::default`] (the
/// production deny-list) so call sites that do not need a custom config
/// remain unchanged.
/// What: Delegates to [`extract_triples_with_config`] with a default config.
/// Test: All existing tests call this helper and implicitly exercise the default
/// deny-list path.
pub fn extract_triples(input: &ExtractInput<'_>) -> Vec<Triple> {
    extract_triples_with_config(input, &KgExtractConfig::default())
}

/// Run the deterministic heuristic extractor.
///
/// Why: Single entry point so `memory_remember`, `memory_note`, and the
/// back-fill CLI all share the same logic. Pure function — no I/O, no async —
/// so it can be unit-tested cheaply. Accepts a [`KgExtractConfig`] so callers
/// can override the deny-list without touching the function signature.
/// What: First checks whether any of the drawer's tags appear in
/// `config.deny_tags` (case-insensitive); when a match is found the function
/// returns immediately with an empty vec and logs a debug message. Otherwise
/// walks `tags`, content tokens, and a small pattern list to emit `Triple`s;
/// deduplicates so the same `(subject, predicate, object)` never appears twice
/// in a single pass.
/// Test: `extract_triples_skips_denied_tags`, `extract_triples_emits_tag_triples`,
/// plus all other tests in this file.
pub fn extract_triples_with_config(
    input: &ExtractInput<'_>,
    config: &KgExtractConfig<'_>,
) -> Vec<Triple> {
    // Deny-list check: if any tag on this drawer is in the deny set, skip
    // extraction entirely. The check is case-insensitive to tolerate mixed-
    // case tags from different clients.
    let denied = input.tags.iter().any(|t| {
        let lower = t.trim().to_lowercase();
        config.deny_tags.contains(&lower.as_str())
    });
    if denied {
        tracing::debug!(
            drawer_id = %input.drawer_id,
            tags = ?input.tags,
            "kg_extract: skipping drawer — tag matches deny-list"
        );
        return Vec::new();
    }
    let now = Utc::now();
    let subject = drawer_subject(input.drawer_id);
    let mut out: Vec<Triple> = Vec::new();
    let mut seen: HashSet<(String, String, String)> = HashSet::new();

    let push = |out: &mut Vec<Triple>,
                seen: &mut HashSet<(String, String, String)>,
                s: String,
                p: String,
                o: String| {
        let key = (s.clone(), p.clone(), o.clone());
        if seen.insert(key) {
            out.push(Triple {
                subject: s,
                predicate: p,
                object: o,
                valid_from: now,
                valid_to: None,
                confidence: AUTO_CONFIDENCE,
                provenance: Some(AUTO_PROVENANCE.to_string()),
            });
        }
    };

    // Tag membership — each tag becomes its own subject so multiple tags on
    // the same drawer don't collide under the "one active triple per
    // (s, p)" invariant. Edge direction is `tag:<t> tags drawer:<id>` so the
    // graph clusters drawers under their shared tag nodes.
    for tag in input.tags {
        let clean = tag.trim();
        if clean.is_empty() {
            continue;
        }
        push(
            &mut out,
            &mut seen,
            format!("{TAG_SUBJECT_PREFIX}{}", clean.to_lowercase()),
            "tags".to_string(),
            subject.clone(),
        );
    }

    // Room membership — `room:<r> contains drawer:<id>` for the same reason
    // (multiple drawers per room must coexist).
    if let Some(room) = input.room {
        let clean = room.trim();
        if !clean.is_empty() {
            push(
                &mut out,
                &mut seen,
                format!("{ROOM_SUBJECT_PREFIX}{clean}"),
                "contains".to_string(),
                subject.clone(),
            );
        }
    }

    // Hashtag-style mentions — `topic:<term> mentioned-in drawer:<id>` so
    // multiple terms per drawer can coexist as distinct active edges.
    for term in extract_hashtags(input.content) {
        push(
            &mut out,
            &mut seen,
            format!("{TOPIC_SUBJECT_PREFIX}{term}"),
            "mentioned-in".to_string(),
            subject.clone(),
        );
    }

    // Simple natural-language patterns. Each yields a free-form
    // `<subject> <predicate> <object>` triple anchored to entities found in
    // the content (not the drawer subject), so the graph develops topical
    // edges over time.
    for (s, p, o) in extract_patterns(input.content, &config.pos) {
        push(&mut out, &mut seen, s, p, o);
    }

    out
}

/// Pull `#hashtag`-style tokens out of free-form content.
///
/// Why: Hashtags are a cheap, intentional signal — when a user writes `#rust`
/// or `#design-doc` we should record the mention so the graph picks it up.
/// What: Walks the string, captures runs of `[a-zA-Z0-9_-]` following a `#`,
/// lower-cases and deduplicates. Skips empty captures (a lone `#`).
/// Test: `extract_triples_emits_hashtag_mentions`.
fn extract_hashtags(content: &str) -> Vec<String> {
    let mut out: Vec<String> = Vec::new();
    let mut seen: HashSet<String> = HashSet::new();
    let mut iter = content.char_indices().peekable();
    while let Some((_, c)) = iter.next() {
        if c != '#' {
            continue;
        }
        let mut term = String::new();
        while let Some(&(_, nc)) = iter.peek() {
            if nc.is_ascii_alphanumeric() || nc == '_' || nc == '-' {
                term.push(nc.to_ascii_lowercase());
                iter.next();
            } else {
                break;
            }
        }
        if term.is_empty() {
            continue;
        }
        if seen.insert(term.clone()) {
            out.push(term);
        }
    }
    out
}

/// Pattern dictionary used by `extract_patterns`.
///
/// Why: A small, predictable set of (predicate, marker phrases) keeps the
/// extractor explicable and deterministic. Each entry maps a predicate to one
/// or more space-padded marker phrases; when the marker appears in the lower-
/// cased content we split on it and read the entity tokens immediately to
/// each side.
/// What: A static slice of `(predicate, &[marker, ...])`. Markers must be
/// lower-case and surrounded by whatever whitespace the input has — we add
/// the padding ourselves.
/// Test: `extract_triples_extracts_is_a_pattern`.
const PATTERN_TABLE: &[(&str, &[&str])] = &[
    ("is-a", &[" is a ", " is an "]),
    ("works-at", &[" works at "]),
    ("uses", &[" uses ", " using "]),
    ("depends-on", &[" depends on ", " requires "]),
];

/// Function words that can never be a KG entity.
///
/// Why: #4678 — a [`PATTERN_TABLE`] marker hit says a marker phrase appeared,
/// not that the tokens either side of it name anything. "calling them is a
/// no-op" put `them --is-a--> no-op` into the live palace. Grouped by part of
/// speech because the boundary is grammatical, not statistical: these are the
/// closed word classes English never coins new members of, so the list is
/// finite and stable rather than a frequency cut-off that needs re-tuning.
/// What: a flat lower-case slice, matched exactly against a normalised token
/// by [`is_stop_token`]. Membership is checked with a linear scan — it runs at
/// most twice per pattern hit and at most four hits exist per drawer.
/// Test: `stopwords_are_unique`, `is_stop_token_rejects_every_stopword`.
const STOPWORDS: &[&str] = &[
    // Articles and determiners.
    "a",
    "an",
    "the",
    "this",
    "that",
    "these",
    "those",
    "some",
    "any",
    "each",
    "every",
    "all",
    "both",
    "either",
    "neither",
    "another",
    "such",
    "same",
    "no",
    "none",
    // Pronouns.
    "i",
    "me",
    "my",
    "mine",
    "myself",
    "we",
    "us",
    "our",
    "ours",
    "ourselves",
    "you",
    "your",
    "yours",
    "yourself",
    "he",
    "him",
    "his",
    "himself",
    "she",
    "her",
    "hers",
    "herself",
    "it",
    "its",
    "itself",
    "they",
    "them",
    "their",
    "theirs",
    "themselves",
    "who",
    "whom",
    "whose",
    "which",
    "what",
    "one",
    "ones",
    "someone",
    "something",
    "anyone",
    "anything",
    "everyone",
    "everything",
    "nobody",
    "nothing",
    "others",
    // Prepositions.
    "of",
    "in",
    "on",
    "at",
    "to",
    "for",
    "with",
    "by",
    "from",
    "about",
    "into",
    "onto",
    "over",
    "under",
    "below",
    "above",
    "between",
    "among",
    "through",
    "during",
    "before",
    "after",
    "across",
    "against",
    "within",
    "without",
    "upon",
    "per",
    "via",
    "than",
    "toward",
    "towards",
    "off",
    "out",
    "up",
    "down",
    "near",
    "around",
    "behind",
    "beyond",
    "beside",
    "along",
    "past",
    "throughout",
    // #5399: WordNet lists these two as nouns ("the inside of the box"), so
    // without the closed-class list they pass the POS check and continue a
    // noun phrase they actually close.
    "inside",
    "outside",
    // Conjunctions and subordinators.
    "and",
    "or",
    "but",
    "nor",
    "so",
    "yet",
    "if",
    "then",
    "else",
    "because",
    "while",
    "when",
    "where",
    "whether",
    "though",
    "although",
    "since",
    "as",
    "unless",
    "until",
    "whereas",
    // Copulas, auxiliaries and modals.
    "is",
    "are",
    "was",
    "were",
    "be",
    "been",
    "being",
    "am",
    "do",
    "does",
    "did",
    "done",
    "doing",
    "has",
    "have",
    "had",
    "having",
    "can",
    "could",
    "will",
    "would",
    "shall",
    "should",
    "may",
    "might",
    "must",
    "ought",
    "need",
    "needs",
    "let",
    "lets",
    "get",
    "gets",
    "got",
    "gotten",
    // Degree, negation and discourse adverbs — closed-class fillers that sit
    // next to a marker often and name nothing.
    "not",
    "only",
    "just",
    "also",
    "very",
    "too",
    "still",
    "already",
    "always",
    "never",
    "often",
    "again",
    "here",
    "there",
    "now",
    "actually",
    "really",
    "simply",
    "merely",
    "quite",
    "rather",
    "even",
    "ever",
    "once",
    "yes",
    "well",
    "much",
    "many",
    "more",
    "most",
    "less",
    "least",
    "few",
    "several",
    "enough",
    "almost",
    "perhaps",
    "maybe",
    "instead",
    "otherwise",
    "hence",
    "therefore",
    "however",
    "thus",
    "moreover",
];

/// Minimum character length for a token to be accepted as an entity.
///
/// Why: one- and two-character tokens are overwhelmingly punctuation debris,
/// initials, or list markers rather than entities.
/// What: `3`. Tokens shorter than this are rejected unless they appear in
/// [`SHORT_ENTITY_ALLOWLIST`].
/// Test: `extract_triples_rejects_short_token_off_allowlist`.
pub const MIN_ENTITY_TOKEN_LEN: usize = 3;

/// Short tokens that are real entities in this workspace's subject matter.
///
/// Why: [`MIN_ENTITY_TOKEN_LEN`] would otherwise reject the languages, crate
/// aliases, and infrastructure abbreviations these palaces actually discuss —
/// a precision filter that silently drops `Go` and `C` costs recall for
/// nothing. The crate aliases (`tm`, `ts`, `tc`, `ta`) come from this repo's
/// own abbreviation table.
/// What: lower-case tokens of fewer than [`MIN_ENTITY_TOKEN_LEN`] characters
/// that bypass the length floor. The stopword check still applies first, so an
/// entry here cannot resurrect a function word.
/// Test: `extract_triples_keeps_allowlisted_go`,
/// `extract_triples_keeps_allowlisted_c`,
/// `short_entity_allowlist_entries_survive_the_length_floor`.
pub const SHORT_ENTITY_ALLOWLIST: &[&str] = &[
    // Languages and language-shaped tokens.
    "go", "c", "c#", "js", "ts", "py", "ml", // Domains and infrastructure.
    "ai", "kg", "db", "ui", "os", "io", "vm", "ip", // Process and workspace aliases.
    "pr", "ci", "qa", "pm", "tm", "tc", "ta",
];

/// Punctuation stripped from a token's edges before it is classified.
///
/// Why: `last_token` / `first_token` only strip a trailing run, so a token can
/// still arrive as `("the` or `` `redb` ``. Comparing that against
/// [`STOPWORDS`] would miss, and the filter would leak exactly the tokens it
/// exists to catch.
/// What: the edge characters [`is_stop_token`] trims before matching. Interior
/// characters are untouched, so `no-op` and `c#` survive intact.
/// Test: `is_stop_token_normalises_surrounding_punctuation`.
const TOKEN_EDGE_PUNCT: &[char] = &[
    '(', ')', '[', ']', '{', '}', '<', '>', '"', '\'', '`', ',', '.', ';', ':', '!', '?', '*', '_',
    '-', '/', '\\', '|', '—', '–', '…', '“', '”', '‘', '’',
];

/// Whether `tok` must be refused as a KG entity.
///
/// Why: #4678 — the pattern pass treated any whitespace-delimited token beside
/// a marker as an entity, which is how `them --is-a--> no-op`,
/// `exhaustiveness --is-a--> hard`, and `squash --is-a--> ancestor` reached the
/// live graph. This is the single gate both the extractor and the
/// `--purge-stale-subjects` back-fill consult, so forward extraction and
/// historical clean-up can never disagree about what counts as garbage.
/// What: normalises `tok` through [`clean_token`], lower-cases it, and rejects
/// it when the result is empty, appears in [`STOPWORDS`], or is shorter than
/// [`MIN_ENTITY_TOKEN_LEN`] without being in [`SHORT_ENTITY_ALLOWLIST`].
/// Length is counted in `char`s, not bytes, so a multi-byte token is not
/// mis-measured. Purely lexical: it judges the token alone and knows nothing of
/// the sentence, so it cannot catch a triple whose tokens are both ordinary
/// words (see #4678 for the residue).
///
/// The normalisation step is NOT redundant with `clean_token`'s use in
/// `first_token` / `last_token`. Those clean tokens on the way IN, which only
/// helps content extracted from now on. The `--purge-stale-subjects` path calls
/// this on subjects read back out of redb, and those were written by the old
/// extractor with the punctuation already welded on — normalising here is what
/// lets a stored `("the` be recognised as the stopword it is.
/// Test: `is_stop_token_rejects_every_stopword`,
/// `is_stop_token_accepts_ordinary_entities`,
/// `is_stop_token_normalises_surrounding_punctuation`,
/// `purge_selects_a_legacy_subject_with_welded_punctuation`.
pub fn is_stop_token(tok: &str) -> bool {
    let norm = clean_token(tok).to_lowercase();
    if norm.is_empty() {
        return true;
    }
    if STOPWORDS.contains(&norm.as_str()) {
        return true;
    }
    norm.chars().count() < MIN_ENTITY_TOKEN_LEN && !SHORT_ENTITY_ALLOWLIST.contains(&norm.as_str())
}

/// The cleaned spelling a stored entity term must be merged onto, or `None`
/// when the term is already canonical.
///
/// Why: #4678's edge trim fixed extraction going forward and split every
/// pre-fix entity in two — a palace holds both `` `redb` `` and `redb`, and
/// each `kg-rebuild` over the same drawer content widens the gap (#5401).
/// Deciding the canonical spelling next to the filter that produced the split
/// is what stops the merge pass from inventing a second normalisation that can
/// drift away from the extractor's.
/// What: `Some(cleaned)` when [`clean_token`] changes `term`; `None` when it
/// does not, or when [`is_stop_token`] rejects the term. That second arm is
/// what keeps this disjoint from `--purge-stale-subjects`: a punctuated
/// stopword such as `("the` is garbage the purge deletes, never a twin some
/// real node should absorb.
///
/// The merge therefore inherits [`clean_token`]'s normalisation whole,
/// dotfiles included: `.env` names `env`, `--verbose` names `verbose`, and the
/// leading punctuation that IS the identity is collapsed with the decorative
/// kind. That is the point of reusing `clean_token` — the extractor already
/// emits `env` for content saying `.env`, so the merge aligns stored nodes
/// with current extraction instead of preserving a spelling nothing produces
/// any more. `close_active_row` keeps the punctuated row under its `hist:`
/// key, so the original spelling is not lost.
/// Test: `canonical_entity_names_the_cleaned_twin`.
pub fn canonical_entity(term: &str) -> Option<&str> {
    // #5401: the merge pass and the extractor must agree on one spelling.
    if is_stop_token(term) {
        return None;
    }
    let cleaned = clean_token(term);
    (cleaned != term).then_some(cleaned)
}

/// Apply the pattern table to a single content blob.
///
/// Why: Keeps the matching loop out of `extract_triples` so the dispatcher
/// stays readable.
/// What: For every `(predicate, markers)` row, scan every marker against the
/// lower-cased content; on the first hit emit `(left_token, predicate,
/// right_token)` and move on to the next predicate. Only the first hit per
/// predicate is taken to avoid combinatorial output on long texts. A hit whose
/// subject or object fails [`is_stop_token`] emits nothing and still consumes
/// the predicate's turn, so one rejected hit never cascades into a scan for a
/// second, lower-quality match later in the blob.
/// Test: `extract_triples_extracts_is_a_pattern`,
/// `extract_triples_rejects_pronoun_subject`,
/// `extract_triples_rejects_stopword_object`.
fn extract_patterns(content: &str, pos: &WordNetPos) -> Vec<(String, String, String)> {
    let lower = content.to_lowercase();
    let mut out: Vec<(String, String, String)> = Vec::new();
    for (predicate, markers) in PATTERN_TABLE {
        for marker in *markers {
            if let Some(idx) = lower.find(marker) {
                // #5399: production hands this whole multi-line drawer bodies
                // (`auto_extract_and_assert`, `kg_rebuild`), so an unbounded
                // walk joins two unrelated sentences across a line break. A
                // newline closes the phrase exactly as a period does.
                let line_start = lower[..idx].rfind('\n').map_or(0, |p| p + 1);
                let left = lower[line_start..idx].trim();
                let right_start = idx + marker.len();
                let line_end = lower[right_start..]
                    .find('\n')
                    .map_or(lower.len(), |p| right_start + p);
                let right = lower[right_start..line_end].trim();
                // #4678: a marker hit is not evidence of two entities — reject
                // the whole triple when either side is a function word or too
                // short, never half of it. #5399: and take the HEAD of each
                // noun phrase rather than the token nearest the marker.
                if let (Some(subject_tok), Some(object_tok)) =
                    (select_subject(left, pos), select_object(right, pos))
                {
                    out.push((subject_tok, (*predicate).to_string(), object_tok));
                }
                break;
            }
        }
    }
    out
}

/// Prepositions that continue a noun phrase rather than closing it.
///
/// Why: #5399 — `ancestor of origin/main` and `member of the process group`
/// are single noun phrases, so grabbing `ancestor` or `member` alone truncates
/// a relation into a bogus type. The stopping token is what reveals it. The set
/// is deliberately just `of`: it is the genitive linker and is almost always
/// NP-internal, whereas every other preposition attaches ambiguously —
/// `trusty-memory uses redb for persistence` has `for` closing the object and
/// opening a purpose adjunct on the verb, and that triple must survive.
/// What: matched against the cleaned token immediately after the extracted
/// noun-phrase run; a hit rejects the whole triple.
/// Test: `row2_rejects_ancestor_truncated_before_of`,
/// `row6_rejects_member_of_the_process_group`,
/// `row5_keeps_uses_object_before_a_non_genitive_preposition`.
const NP_CONTINUING_PREPOSITIONS: &[&str] = &["of"];

/// Characters that close a noun phrase when welded to a token's trailing edge.
///
/// Why: `is a parser, and ...` must not walk the run into the next clause.
/// What: consulted on the RAW token, through [`ends_noun_phrase`], before
/// [`clean_token`] strips it.
/// Test: `stops_the_noun_phrase_run_at_a_comma`,
/// `stops_the_run_at_a_terminator_behind_markdown_emphasis`.
const NP_TERMINATING_PUNCT: &[char] = &['.', ',', ';', ':', '!', '?', ')'];

/// Longest noun-phrase run either walk will consider.
///
/// Why: a bound keeps a pathological line (a long unpunctuated list of unknown
/// tokens) from letting the head drift far from the marker. Four covers
/// `[adj] [noun] [noun]` plus one, which is past the length of any real
/// technical compound in drawer prose.
/// Test: `caps_the_noun_phrase_run`.
const NP_RUN_MAX: usize = 4;

/// Whether a raw token's trailing punctuation closes the noun phrase.
///
/// Why: #5399 — the check used to read only the token's LAST character, so
/// `**MCP is a thin proxy.**` (raw token `proxy.**`) ended in `*`, missed the
/// `.`, and let the run walk into the next clause. Markdown emphasis is the
/// normal shape of drawer content, not an edge case, so the terminator is
/// almost always behind one or two more punctuation characters.
/// What: scans the whole trailing punctuation run — every trailing character
/// that [`clean_token`] would strip — and reports whether any of it terminates
/// a sentence or clause. Interior punctuation is not consulted, so `Node.js`
/// and `src/main.rs` still read as one unterminated token.
/// Test: `stops_the_run_at_a_terminator_behind_markdown_emphasis`,
/// `interior_punctuation_does_not_terminate_the_run`.
fn ends_noun_phrase(raw: &str) -> bool {
    raw.trim_end()
        .chars()
        .rev()
        .take_while(|c| TOKEN_EDGE_PUNCT.contains(c))
        .any(|c| NP_TERMINATING_PUNCT.contains(&c))
}

/// Which way a noun-phrase walk moves away from the marker.
///
/// Why: the two sides are mirror images in one respect that matters. Walking
/// RIGHT, a token's trailing `.` closes the phrase we are building, so the
/// token belongs to it. Walking LEFT, that same `.` ended the PREVIOUS
/// sentence, so the token belongs to that one and must not be taken.
/// Test: `stops_the_noun_phrase_run_at_a_comma`,
/// `subject_walk_stops_before_a_previous_sentence`.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum Walk {
    /// Away from the marker into the object phrase.
    Right,
    /// Away from the marker into the subject phrase.
    Left,
}

/// Collect the noun-phrase run adjacent to a marker.
///
/// Why: both sides of a pattern hit need the same walk — "which tokens belong
/// to this phrase" is one question with one answer, and writing it twice is how
/// the two sides drift.
/// What: consumes `toks` in walk order (the caller reverses them for
/// [`Walk::Left`]) and stops at a function word, a verb-only or adverb-only
/// token, sentence punctuation, an unknown token, or [`NP_RUN_MAX`]. Returns
/// the accepted tokens in READING order regardless of direction, whether
/// punctuation closed the phrase, and how many raw tokens were consumed — the
/// caller needs the last two to decide about a following genitive.
/// Test: `caps_the_noun_phrase_run`, `stops_the_noun_phrase_run_at_a_comma`,
/// `subject_walk_stops_before_a_previous_sentence`.
fn noun_phrase_run<'a>(
    toks: &[&'a str],
    pos: &WordNetPos,
    walk: Walk,
) -> (Vec<&'a str>, bool, usize) {
    let mut run: Vec<&str> = Vec::new();
    let mut idx = 0usize;
    let mut terminated = false;
    while idx < toks.len() && run.len() < NP_RUN_MAX {
        let raw = toks[idx];
        let tok = clean_token(raw);
        if tok.is_empty() || is_stop_token(tok) {
            break;
        }
        let closes = ends_noun_phrase(raw);
        // Walking left, the punctuation sits between this token and the phrase
        // we are collecting, so the token is on the far side of the boundary.
        if closes && walk == Walk::Left {
            terminated = true;
            break;
        }
        let mask = pos.mask(tok);
        // A word WordNet knows only as a verb or only as an adverb cannot sit
        // inside a noun phrase, so the phrase ended before it.
        if mask != 0 && mask & (wordnet_pos::NOUN | wordnet_pos::ADJ) == 0 {
            break;
        }
        run.push(tok);
        idx += 1;
        if closes {
            terminated = true;
            break;
        }
        // A name heads its phrase rather than modifying the next word.
        if mask == 0 {
            break;
        }
    }
    if walk == Walk::Left {
        run.reverse();
    }
    (run, terminated, idx)
}

/// Pick the head of a noun-phrase run.
///
/// Why: #5399 — English compounds are right-headed, so the head is the
/// rightmost token of the phrase that can name a thing. An adjective-only token
/// names a property, so it is skipped rather than taken. This RE-WALK is the
/// whole policy: the earlier spike instead REJECTED the triple when the token
/// nearest the marker was adjective-only, and that dropped 304 pairs over 306k
/// lines of repo markdown, 196 of them (64%) with a perfectly good head noun
/// sitting one token further along. See the fixture comment on
/// `row1_rewalks_past_an_adjective_only_modifier` for why the reject rule was
/// wrong on its own terms and not merely expensive.
/// What: `run` is in reading order, so this returns its last non-adjective-only
/// token — or `None` when every token in it is adjective-only (`a robust` names
/// nothing) or the run is empty.
/// Test: `row1_rewalks_past_an_adjective_only_modifier`,
/// `row4_rewalks_past_the_adjective_to_the_head_noun`,
/// `an_all_adjective_subject_yields_no_triple`.
fn phrase_head(run: &[&str], pos: &WordNetPos) -> Option<String> {
    run.iter()
        .rev()
        .find(|t| !pos.is_adjective_only(t))
        .map(|t| (*t).to_string())
}

/// Pick the subject entity out of the text preceding a pattern marker.
///
/// Why: the head of the subject phrase is normally the token immediately before
/// the marker — `a fast parser is a tool` — so this agrees with the old
/// `last_token` in every ordinary case. It differs only when that token cannot
/// head a phrase, where `last_token` emitted it as an entity anyway. That is
/// the #5399 defect (`exhaustiveness --is-a--> hard`) mirrored onto the other
/// side of the marker, so it gets the same re-walk rather than a reject rule.
/// What: walks leftward from the marker through [`noun_phrase_run`] and takes
/// the phrase head. Returns `None` when the phrase has no head — `anything
/// robust is a compiler` names no entity to be the subject of anything.
/// Test: `subject_side_rewalks_past_an_adjective_only_token`,
/// `an_all_adjective_subject_yields_no_triple`,
/// `extract_triples_rejects_pronoun_subject`.
fn select_subject(left: &str, pos: &WordNetPos) -> Option<String> {
    let mut toks: Vec<&str> = left.split_whitespace().collect();
    toks.reverse();
    let (run, _, _) = noun_phrase_run(&toks, pos, Walk::Left);
    phrase_head(&run, pos)
}

/// Pick the object entity out of the text following a pattern marker.
///
/// Why: #5399 — taking one token is wrong three ways at once. It takes the
/// modifier instead of the head (`a fast parser` -> `fast`), it emits a bare
/// property as a type (`a hard requirement` -> `hard`), and it silently
/// truncates a relation into a type (`an ancestor of origin main` ->
/// `ancestor`). All three need to see more than one token and need to know each
/// token's part of speech.
/// What: walks rightward from the marker through [`noun_phrase_run`], rejects
/// the triple when the token that STOPPED the run is in
/// [`NP_CONTINUING_PREPOSITIONS`] (the phrase was not finished, so its head is
/// not in what we collected), and otherwise returns the phrase head. Every
/// WordNet miss widens the accepted set rather than narrowing it, so an unknown
/// crate name is never rejected for being unknown.
/// Test: `row4_rewalks_past_the_adjective_to_the_head_noun`,
/// `row2_rejects_ancestor_truncated_before_of`,
/// `unknown_subject_and_object_both_fail_open`.
fn select_object(right: &str, pos: &WordNetPos) -> Option<String> {
    let toks: Vec<&str> = right.split_whitespace().collect();
    let (run, terminated, consumed) = noun_phrase_run(&toks, pos, Walk::Right);
    if !terminated
        && toks
            .get(consumed)
            .is_some_and(|next| NP_CONTINUING_PREPOSITIONS.contains(&clean_token(next)))
    {
        return None;
    }
    phrase_head(&run, pos)
}

/// Strip surrounding punctuation from one raw token.
///
/// Why: #4678 — `first_token` trimmed a TRAILING run while its doc promised a
/// leading one, and both helpers used a set that omitted the characters drawer
/// content actually wraps names in (backticks, brackets, asterisks). So
/// `` `redb` `` reached the graph verbatim and became a second node for an
/// entity that already had one. Which SIDE of the marker a token sits on says
/// nothing about which edge carries punctuation, so both helpers clean both
/// edges through this one function rather than each guessing.
/// What: trims whitespace, then [`TOKEN_EDGE_PUNCT`] from both ends. Interior
/// characters are untouched, so `no-op`, `c#`, and `src/main.rs` survive whole.
/// Test: `extract_triples_strips_punctuation_from_both_token_edges`.
fn clean_token(raw: &str) -> &str {
    raw.trim().trim_matches(TOKEN_EDGE_PUNCT)
}

#[cfg(test)]
mod tests {
    use super::*;

    fn input_for(content: &str, tags: &[&str], room: Option<&str>) -> (Uuid, Vec<String>) {
        let id = Uuid::new_v4();
        let owned_tags: Vec<String> = tags.iter().map(|s| s.to_string()).collect();
        let _ = content; // silence unused warning if test ignores content
        let _ = room;
        (id, owned_tags)
    }

    /// Why: Tag-derived triples are the lowest-hanging extraction and the
    /// graph view's first signal when no patterns fire. The KG's temporal
    /// model only allows one active triple per `(subject, predicate)`, so
    /// each tag becomes its own subject (`tag:<name>`) with a `tags`
    /// predicate pointing at the drawer.
    /// What: One `tag:<t> tags drawer:<id>` per non-empty tag, plus
    /// `room:<r> contains drawer:<id>` when a room is supplied.
    /// Test: This test.
    #[test]
    fn extract_triples_emits_tag_triples() {
        let (id, tags) = input_for("hello world", &["rust", "design"], Some("Backend"));
        let triples = extract_triples(&ExtractInput {
            drawer_id: id,
            content: "hello world",
            tags: &tags,
            room: Some("Backend"),
        });
        let object = drawer_subject(id);
        assert!(triples
            .iter()
            .any(|t| t.subject == "tag:rust" && t.predicate == "tags" && t.object == object));
        assert!(triples
            .iter()
            .any(|t| t.subject == "tag:design" && t.predicate == "tags" && t.object == object));
        assert!(triples.iter().any(|t| t.subject == "room:Backend"
            && t.predicate == "contains"
            && t.object == object));
    }

    /// Why: Hashtag tokens are a cheap user signal; the extractor must catch
    /// them so the graph picks up topical entities.
    /// What: `#rust` and `#design-doc` both become `topic:<term>
    /// mentioned-in drawer:<id>` triples, lower-cased and deduplicated.
    /// Test: This test.
    #[test]
    fn extract_triples_emits_hashtag_mentions() {
        let (id, tags) = input_for("see #Rust and #design-doc and #rust again", &[], None);
        let triples = extract_triples(&ExtractInput {
            drawer_id: id,
            content: "see #Rust and #design-doc and #rust again",
            tags: &tags,
            room: None,
        });
        let mention_subjects: Vec<&str> = triples
            .iter()
            .filter(|t| t.predicate == "mentioned-in")
            .map(|t| t.subject.as_str())
            .collect();
        assert!(mention_subjects.contains(&"topic:rust"));
        assert!(mention_subjects.contains(&"topic:design-doc"));
        // Dedupe — `#rust` and `#Rust` collapse.
        assert_eq!(
            mention_subjects
                .iter()
                .filter(|s| **s == "topic:rust")
                .count(),
            1
        );
    }

    /// Why: `is a` is the simplest NL pattern and the most common idiom in
    /// quick notes ("rustc is a compiler").
    /// What: Pattern fires once per content blob; subject and object are the
    /// nouns either side of the marker.
    /// Test: This test.
    #[test]
    fn extract_triples_extracts_is_a_pattern() {
        let (id, _) = input_for("rustc is a compiler for rust", &[], None);
        let triples = extract_triples(&ExtractInput {
            drawer_id: id,
            content: "rustc is a compiler for rust",
            tags: &[],
            room: None,
        });
        assert!(triples
            .iter()
            .any(|t| t.subject == "rustc" && t.predicate == "is-a" && t.object == "compiler"));
    }

    /// Why: Confidence and provenance are guard-rails — extracted triples
    /// must be recognisable and over-ridable.
    /// What: Every triple carries `provenance = Some("auto:remember")` and
    /// `confidence == AUTO_CONFIDENCE`.
    /// Test: This test.
    #[test]
    fn extract_triples_stamps_provenance() {
        let (id, tags) = input_for("anything", &["x"], None);
        let triples = extract_triples(&ExtractInput {
            drawer_id: id,
            content: "anything",
            tags: &tags,
            room: None,
        });
        assert!(!triples.is_empty());
        for t in &triples {
            assert_eq!(t.provenance.as_deref(), Some(AUTO_PROVENANCE));
            assert!((t.confidence - AUTO_CONFIDENCE).abs() < f32::EPSILON);
        }
    }

    /// Why: Reduced confidence is the contract a manual `kg_assert` of the
    /// same `(subject, predicate)` needs in order to "win" against the
    /// auto-extracted edge.
    /// What: Every triple carries `confidence == AUTO_CONFIDENCE` (currently
    /// 0.6); the constant is asserted to stay strictly below 1.0 so manual
    /// asserts always rank higher.
    /// Test: This test.
    #[test]
    #[allow(clippy::assertions_on_constants)]
    fn extract_triples_uses_reduced_confidence() {
        // Why: both bounds are static facts about the AUTO_CONFIDENCE
        // constant; the assertion is documentation for future tweakers.
        assert!(AUTO_CONFIDENCE < 1.0);
        assert!(AUTO_CONFIDENCE > 0.0);
    }

    /// Why: Empty / whitespace-only content must not panic or emit garbage.
    /// What: No tags, no room, no content → empty vec.
    /// Test: This test.
    #[test]
    fn extract_triples_never_panics_on_empty_input() {
        let id = Uuid::new_v4();
        let triples = extract_triples(&ExtractInput {
            drawer_id: id,
            content: "",
            tags: &[],
            room: None,
        });
        assert!(triples.is_empty());
    }

    /// Why: Edge-case test — content with no patterns but tags should still
    /// produce the tag triples (the graph view's primary signal).
    /// What: Single tag, no room, prose with no pattern hits → exactly one
    /// triple shaped as `tag:meeting tags drawer:<id>`.
    /// Test: This test.
    #[test]
    fn extract_triples_tags_only_path() {
        let id = Uuid::new_v4();
        let tags = vec!["meeting".to_string()];
        let triples = extract_triples(&ExtractInput {
            drawer_id: id,
            content: "Discussed roadmap.",
            tags: &tags,
            room: None,
        });
        assert_eq!(triples.len(), 1);
        assert_eq!(triples[0].subject, "tag:meeting");
        assert_eq!(triples[0].predicate, "tags");
        assert_eq!(triples[0].object, drawer_subject(id));
    }

    /// Why: Drawers tagged with deny-listed labels (test fixtures, QA scaffolding)
    /// must not pollute the KG with non-factual content.
    /// What: A drawer with the `test` tag must produce zero triples even when
    /// it also has a room and content with extractable patterns.
    /// Test: This test.
    #[test]
    fn extract_triples_skips_denied_tags() {
        let id = Uuid::new_v4();
        let tags = vec!["test".to_string(), "rust".to_string()];
        let triples = extract_triples(&ExtractInput {
            drawer_id: id,
            content: "rustc is a compiler",
            tags: &tags,
            room: Some("Backend"),
        });
        assert!(
            triples.is_empty(),
            "a drawer with a deny-list tag must produce zero triples, got {triples:?}"
        );
    }

    /// Why: Deny-list matching is case-insensitive so `TEST` and `Test` are
    /// blocked the same as `test`.
    /// What: A drawer tagged `FIXTURE` (upper-case) must still produce zero
    /// triples.
    /// Test: This test.
    #[test]
    fn extract_triples_deny_list_is_case_insensitive() {
        let id = Uuid::new_v4();
        let tags = vec!["FIXTURE".to_string()];
        let triples = extract_triples(&ExtractInput {
            drawer_id: id,
            content: "some content",
            tags: &tags,
            room: None,
        });
        assert!(
            triples.is_empty(),
            "upper-cased deny tag must still be blocked"
        );
    }

    /// Collect only the triples produced by the [`PATTERN_TABLE`] pass.
    ///
    /// Why: every assertion about extraction precision is about the pattern
    /// pass; the tag / room / hashtag passes always fire and would otherwise
    /// mask a "zero triples" assertion.
    /// What: filters by predicate against [`PATTERN_TABLE`].
    /// Test: used by every precision test below.
    fn pattern_triples(triples: &[Triple]) -> Vec<(String, String, String)> {
        let predicates: Vec<&str> = PATTERN_TABLE.iter().map(|(p, _)| *p).collect();
        triples
            .iter()
            .filter(|t| predicates.contains(&t.predicate.as_str()))
            .map(|t| (t.subject.clone(), t.predicate.clone(), t.object.clone()))
            .collect()
    }

    /// Run the extractor over bare content with no tags and no room.
    ///
    /// Why: the precision fixtures care only about content; tags and rooms
    /// would add noise triples to every assertion.
    /// What: builds an `ExtractInput` with empty tags and no room.
    /// Test: used by every precision test below.
    fn patterns_for(content: &str) -> Vec<(String, String, String)> {
        let triples = extract_triples(&ExtractInput {
            drawer_id: Uuid::new_v4(),
            content,
            tags: &[],
            room: None,
        });
        pattern_triples(&triples)
    }

    /// Why: `them --is-a--> no-op` is live in the real trusty-tools palace
    /// (asserted 2026-08-04). A marker hit is not evidence that the tokens
    /// either side of it are entities; a pronoun never is one.
    /// What: "calling them is a no-op ..." must yield zero pattern triples.
    /// Test: This test.
    #[test]
    fn extract_triples_rejects_pronoun_subject() {
        let got = patterns_for("calling them is a no-op when the flag is off");
        assert!(
            got.is_empty(),
            "pronoun subject must reject the whole triple, got {got:?}"
        );
    }

    /// Why: a stopword can land on either side of a marker, so filtering only
    /// the subject would still admit `libpq --depends-on--> the`.
    /// What: "libpq depends on the license" must yield zero pattern triples —
    /// the triple is rejected whole, never truncated to a half-triple.
    /// Test: This test.
    #[test]
    fn extract_triples_rejects_stopword_object() {
        let got = patterns_for("libpq depends on the license header");
        assert!(
            got.is_empty(),
            "stopword object must reject the whole triple, got {got:?}"
        );
    }

    /// Why: a one- or two-character token is almost never an entity, and the
    /// extractor had no length floor at all.
    /// What: "x is a thing" must yield zero pattern triples.
    /// Test: This test.
    #[test]
    fn extract_triples_rejects_short_token_off_allowlist() {
        let got = patterns_for("x is a thing worth recording");
        assert!(
            got.is_empty(),
            "short token off the allowlist must be rejected, got {got:?}"
        );
    }

    /// Why: the length floor must not swallow the short names this repo
    /// genuinely discusses — without the allowlist, `Go` and `C` would be
    /// rejected as noise and the filter would cost real recall.
    /// What: "Go is a language" still extracts `go --is-a--> language`.
    /// Test: This test.
    #[test]
    fn extract_triples_keeps_allowlisted_go() {
        let got = patterns_for("Go is a language with green threads");
        assert!(
            got.contains(&("go".into(), "is-a".into(), "language".into())),
            "allowlisted `Go` must survive the length floor, got {got:?}"
        );
    }

    /// Why: `C` is the single-character case — the length floor's worst
    /// false positive if the allowlist is not consulted.
    /// What: "C is a language" still extracts `c --is-a--> language`.
    /// Test: This test.
    #[test]
    fn extract_triples_keeps_allowlisted_c() {
        let got = patterns_for("C is a language without a runtime");
        assert!(
            got.contains(&("c".into(), "is-a".into(), "language".into())),
            "allowlisted `C` must survive the length floor, got {got:?}"
        );
    }

    /// Why: the filter must not cost recall on ordinary, well-formed content.
    /// A rejection filter that also rejects the good cases is a regression,
    /// so every predicate in [`PATTERN_TABLE`] keeps a worked example.
    /// What: table-driven — each row is `(content, expected triple)` and must
    /// appear in the extracted pattern set.
    /// Test: This test.
    #[test]
    fn extract_triples_keeps_real_entities_across_all_predicates() {
        let cases: &[(&str, (&str, &str, &str))] = &[
            (
                "tokio is an executor for async rust",
                ("tokio", "is-a", "executor"),
            ),
            (
                "alice works at initech today",
                ("alice", "works-at", "initech"),
            ),
            (
                "trusty-memory uses redb for persistence",
                ("trusty-memory", "uses", "redb"),
            ),
            (
                "trusty-search depends on trusty-common for shared helpers",
                ("trusty-search", "depends-on", "trusty-common"),
            ),
            (
                "the embedder requires onnxruntime at startup",
                ("embedder", "depends-on", "onnxruntime"),
            ),
        ];
        for (content, (s, p, o)) in cases {
            let got = patterns_for(content);
            let want = ((*s).to_string(), (*p).to_string(), (*o).to_string());
            assert!(
                got.contains(&want),
                "content {content:?} must still extract {want:?}, got {got:?}"
            );
        }
    }

    /// Why: `first_token` trimmed only a TRAILING run while its doc promised it
    /// stripped the leading one, so an entity quoted the way drawer content
    /// actually quotes things — markdown backticks, brackets, an opening
    /// quote — entered the graph with the punctuation welded on: `` `redb ``
    /// rather than `redb`. Two spellings of one entity are two nodes that never
    /// join up.
    /// What: pins the emitted strings. Both helpers strip punctuation from BOTH
    /// edges, so the subject side (`last_token`) is covered too, and no emitted
    /// token may retain an edge character.
    /// Test: This test.
    #[test]
    fn extract_triples_strips_punctuation_from_both_token_edges() {
        let cases: &[(&str, (&str, &str, &str))] = &[
            // Object side, markdown backticks — the common shape in drawers.
            (
                "trusty-memory uses `redb` for persistence",
                ("trusty-memory", "uses", "redb"),
            ),
            // Object side, parenthesised.
            (
                "the daemon uses (tantivy) for search",
                ("daemon", "uses", "tantivy"),
            ),
            // Subject side, opening quote with no closing one before the marker.
            (
                "he said \"rustc is a compiler for rust",
                ("rustc", "is-a", "compiler"),
            ),
            // Object side, bold markdown.
            (
                "trusty-search depends on **trusty-common** for helpers",
                ("trusty-search", "depends-on", "trusty-common"),
            ),
        ];
        for (content, (s, p, o)) in cases {
            let got = patterns_for(content);
            let want = ((*s).to_string(), (*p).to_string(), (*o).to_string());
            assert!(
                got.contains(&want),
                "content {content:?} must emit {want:?}, got {got:?}"
            );
            for (subject, _, object) in &got {
                for tok in [subject, object] {
                    assert!(
                        !tok.starts_with(TOKEN_EDGE_PUNCT) && !tok.ends_with(TOKEN_EDGE_PUNCT),
                        "emitted token {tok:?} still carries edge punctuation"
                    );
                }
            }
        }
    }

    /// Why: a duplicated entry is dead weight and a sign the categories drifted
    /// apart during editing.
    /// What: every [`STOPWORDS`] entry appears exactly once and is already
    /// lower-case, which is what [`is_stop_token`] compares against.
    /// Test: This test.
    #[test]
    fn stopwords_are_unique() {
        let mut seen: HashSet<&str> = HashSet::new();
        for w in STOPWORDS {
            assert!(seen.insert(w), "duplicate stopword {w:?}");
            assert_eq!(*w, w.to_lowercase(), "stopword {w:?} must be lower-case");
        }
    }

    /// Why: the rejection contract is only as good as the list behind it.
    /// What: every [`STOPWORDS`] entry is rejected, in its own case and
    /// upper-cased, since content reaches the filter lower-cased but the purge
    /// path reads subjects straight out of the store.
    /// Test: This test.
    #[test]
    fn is_stop_token_rejects_every_stopword() {
        for w in STOPWORDS {
            assert!(is_stop_token(w), "{w:?} must be rejected");
            assert!(
                is_stop_token(&w.to_uppercase()),
                "{w:?} must be rejected case-insensitively"
            );
        }
    }

    /// Why: an over-broad filter costs recall, which is the failure mode that
    /// would make this change a net loss.
    /// What: ordinary multi-character entity names are accepted.
    /// Test: This test.
    #[test]
    fn is_stop_token_accepts_ordinary_entities() {
        for tok in [
            "rustc",
            "compiler",
            "trusty-memory",
            "redb",
            "no-op",
            "onnxruntime",
            "initech",
        ] {
            assert!(!is_stop_token(tok), "{tok:?} must be accepted");
        }
    }

    /// Why: `first_token` / `last_token` strip only a trailing run, so a token
    /// can still carry an opening bracket or backtick. Comparing that raw
    /// against the list would miss the stopword it is meant to catch.
    /// What: edge punctuation is stripped before matching, interior characters
    /// are not, and a token that is nothing but punctuation is rejected.
    /// Test: This test.
    #[test]
    fn is_stop_token_normalises_surrounding_punctuation() {
        assert!(
            is_stop_token("(the"),
            "leading bracket must not hide a stopword"
        );
        assert!(is_stop_token("`it`"), "backticks must not hide a stopword");
        assert!(
            is_stop_token("---"),
            "punctuation-only token must be rejected"
        );
        assert!(
            is_stop_token("   "),
            "whitespace-only token must be rejected"
        );
        assert!(
            !is_stop_token("`redb`"),
            "interior name must survive trimming"
        );
        assert!(
            !is_stop_token("no-op"),
            "interior hyphen must survive trimming"
        );
    }

    /// Why: #5401's merge asks this function which node a stored term belongs
    /// under, and its three answers are the pass's whole selection rule.
    /// What: a punctuated real entity names its cleaned twin; an already-clean
    /// term and a punctuated stopword both name nothing — the second because it
    /// is `--purge-stale-subjects`'s to delete, not this pass's to re-point. A
    /// name whose leading punctuation IS the identity — `.env`, `--verbose` —
    /// names the same twin as decorative punctuation does, because the merge
    /// inherits [`clean_token`] rather than re-deciding.
    /// Test: This test.
    #[test]
    fn canonical_entity_names_the_cleaned_twin() {
        assert_eq!(canonical_entity("`redb`"), Some("redb"));
        assert_eq!(canonical_entity("(sled)"), Some("sled"));
        assert_eq!(canonical_entity("*trusty-memory*"), Some("trusty-memory"));
        assert_eq!(
            canonical_entity("src/main.rs,"),
            Some("src/main.rs"),
            "only edge punctuation moves; the interior path must survive"
        );

        assert_eq!(canonical_entity("redb"), None, "a clean term is canonical");
        assert_eq!(canonical_entity("no-op"), None);

        assert_eq!(
            canonical_entity("(\"the"),
            None,
            "a punctuated stopword belongs to the purge, not the merge"
        );
        assert_eq!(canonical_entity("`it`"), None);
        assert_eq!(canonical_entity("---"), None);

        // The dotfile class: the leading punctuation is part of the name, and
        // it is collapsed anyway. Deliberate — today's extractor already emits
        // `env` for a drawer that says `.env`, so the merge aligns the stored
        // node with what extraction now produces rather than inventing a
        // second normalisation that could drift from it.
        assert_eq!(canonical_entity(".env"), Some("env"));
        assert_eq!(canonical_entity(".git"), Some("git"));
        assert_eq!(canonical_entity(".gitignore"), Some("gitignore"));
        assert_eq!(canonical_entity("_private"), Some("private"));
        assert_eq!(canonical_entity("--verbose"), Some("verbose"));
    }

    /// Why: the allowlist is the length floor's escape hatch; if an entry stops
    /// working the floor silently starts eating real entities.
    /// What: every [`SHORT_ENTITY_ALLOWLIST`] entry is accepted, and no entry
    /// collides with [`STOPWORDS`] (which is checked first and would win).
    /// Test: This test.
    #[test]
    fn short_entity_allowlist_entries_survive_the_length_floor() {
        for tok in SHORT_ENTITY_ALLOWLIST {
            assert!(
                !STOPWORDS.contains(tok),
                "{tok:?} is on both lists; the stopword check runs first and wins"
            );
            assert!(!is_stop_token(tok), "allowlisted {tok:?} must be accepted");
        }
    }

    /// Why: #4678 named three live regressions and could only reach the first.
    /// Its filter is lexical — it judges one token at a time — and
    /// `exhaustiveness`, `hard`, `squash` and `ancestor` are all ordinary
    /// content words, indistinguishable token-wise from the `rustc --is-a-->
    /// compiler` the extractor is supposed to keep. So #4678 shipped this test
    /// ASSERTING the residue, under the name
    /// `lexical_filter_does_not_reach_the_two_content_word_regressions`, as the
    /// standing record of an open gate.
    ///
    /// #5399 closed that gate with WordNet head-noun selection, so the
    /// assertion is INVERTED rather than deleted and the name is kept
    /// recognisable — `git log -S` on either name finds both states. The
    /// lexical half is unchanged and still passes: it is what proves the fix
    /// landed in the POS pass and not by quietly widening the stopword list,
    /// which would have broken `rustc --is-a--> compiler` too.
    /// What: asserts the four words still survive the lexical filter, AND that
    /// neither #4678 residue triple is produced any more.
    /// Test: This test.
    #[test]
    fn lexical_filter_still_cannot_reach_the_two_content_word_regressions_but_wordnet_does() {
        assert!(
            !is_stop_token("exhaustiveness")
                && !is_stop_token("hard")
                && !is_stop_token("squash")
                && !is_stop_token("ancestor"),
            "these are ordinary words; no lexical filter may reject them"
        );
        let exhaustiveness = patterns_for("match exhaustiveness is a hard requirement here");
        assert!(
            !exhaustiveness.contains(&("exhaustiveness".into(), "is-a".into(), "hard".into())),
            "#4678 residue must be gone; got {exhaustiveness:?}"
        );
        let squash = patterns_for("confirm the squash is an ancestor of origin main");
        assert!(
            !squash.contains(&("squash".into(), "is-a".into(), "ancestor".into())),
            "#4678 residue must be gone; got {squash:?}"
        );
    }

    /// Why: An empty deny-list (e.g. in integration tests that want to exercise
    /// extraction regardless of tags) must not suppress any triples.
    /// What: Calling `extract_triples_with_config` with `deny_tags = &[]` on a
    /// drawer tagged `test` must produce the normal tag triple.
    /// Test: This test.
    #[test]
    fn extract_triples_empty_deny_list_passes_through() {
        let id = Uuid::new_v4();
        let tags = vec!["test".to_string()];
        let config = KgExtractConfig {
            deny_tags: &[],
            ..Default::default()
        };
        let triples = extract_triples_with_config(
            &ExtractInput {
                drawer_id: id,
                content: "anything",
                tags: &tags,
                room: None,
            },
            &config,
        );
        // "test" tag should produce a tag triple when the deny-list is empty.
        assert!(
            !triples.is_empty(),
            "empty deny-list must not suppress extraction"
        );
    }
}

/// The #5399 evaluation set for the shipped WordNet policy, plus the cases that
/// probe where it behaves surprisingly.
///
/// Why: these rows decided the bake-off between the WordNet lane and a spaCy
/// lane on identical inputs, so they are the record of what was chosen. The
/// `surprising_*` tests are deliberately included even where they record a
/// WRONG answer — an eval set built only from passing rows tells the next
/// reader nothing about where the approach runs out.
/// What: each test drives `extract_triples` end to end and asserts on the
/// pattern-derived triples only (tag/room/hashtag triples are unrelated).
#[cfg(test)]
mod wordnet_eval {
    use super::*;

    /// Pattern triples only, as `(subject, predicate, object)`.
    fn pattern_triples(content: &str) -> Vec<(String, String, String)> {
        extract_triples(&ExtractInput {
            drawer_id: Uuid::new_v4(),
            content,
            tags: &[],
            room: None,
        })
        .into_iter()
        .filter(|t| PATTERN_TABLE.iter().any(|(p, _)| *p == t.predicate))
        .map(|t| (t.subject, t.predicate, t.object))
        .collect()
    }

    fn assert_none(content: &str) {
        let got = pattern_triples(content);
        assert!(
            got.is_empty(),
            "expected no triple from {content:?}, got {got:?}"
        );
    }

    fn assert_one(content: &str, s: &str, p: &str, o: &str) {
        let got = pattern_triples(content);
        assert_eq!(
            got,
            vec![(s.to_string(), p.to_string(), o.to_string())],
            "wrong extraction from {content:?}"
        );
    }

    // ---- Row 1: adjective-only modifier. `hard` is ADJ|ADV, never NOUN. ----
    //
    // 🔴 THIS ROW'S EXPECTATION WAS CHANGED, AND THE OLD ONE WAS WRONG. The
    // lane-A spike specified "NO triple" here, on the theory that a phrase
    // whose modifier is adjective-only is unsalvageable. That misread the
    // defect. #5399 is about extracting the ADJECTIVE `hard` as though it were
    // a type; it is not about the noun `requirement`, and
    // `exhaustiveness --is-a--> requirement` is a TRUE statement about this
    // sentence. Rejecting the whole triple threw away a good fact to avoid a
    // bad one.
    //
    // The corpus settled it: over 306k lines of repo markdown the reject rule
    // dropped 304 pairs, and 196 of them (64%) had a recoverable head noun one
    // token further along. 43 of 78 ordinary technical adjectives — `robust`,
    // `concurrent`, `embedded`, `distributed`, `immutable`, `scalable` — are
    // adjective-only in WordNet, so the rule fired constantly, and every time
    // it fired it destroyed a usable triple.
    //
    // Do NOT "fix" this back to `assert_none`. Re-walking to the head noun IS
    // the shipped policy; a test asserting no triple here asserts the bug.
    #[test]
    fn row1_rewalks_past_an_adjective_only_modifier() {
        assert_one(
            "match exhaustiveness is a hard requirement here",
            "exhaustiveness",
            "is-a",
            "requirement",
        );
    }

    // ---- Row 2: genitive boundary. The NP is `ancestor of origin main`. ----
    #[test]
    fn row2_rejects_ancestor_truncated_before_of() {
        assert_none("confirm the squash is an ancestor of origin main");
    }

    // ---- Row 3: the baseline good triple must survive untouched. ----
    #[test]
    fn row3_keeps_rustc_is_a_compiler() {
        assert_one("rustc is a compiler", "rustc", "is-a", "compiler");
    }

    // ---- Row 4: head-noun re-walk past a modifier. ----
    #[test]
    fn row4_rewalks_past_the_adjective_to_the_head_noun() {
        assert_one("librs is a fast parser", "librs", "is-a", "parser");
    }

    // ---- Row 5: `for` closes the object; it is a verb adjunct, not a NP. ----
    #[test]
    fn row5_keeps_uses_object_before_a_non_genitive_preposition() {
        assert_one(
            "trusty-memory uses redb for persistence",
            "trusty-memory",
            "uses",
            "redb",
        );
    }

    // ---- Row 6: same genitive shape as row 2, different noun. ----
    #[test]
    fn row6_rejects_member_of_the_process_group() {
        assert_none("the daemon is a member of the process group");
    }

    // ---- Row 7: right-headed compound; the head is the LAST noun. ----
    #[test]
    fn row7_takes_the_head_of_a_noun_noun_compound() {
        assert_one("tantivy is a search library", "tantivy", "is-a", "library");
    }

    // ================== the re-walk, on the cases that drove it =============

    /// The adjectives with no WordNet noun sense. Under the spike's reject rule
    /// every one of these produced nothing; under the shipped policy each
    /// re-walks to its head noun.
    #[test]
    fn adjective_only_modifiers_re_walk_instead_of_destroying_the_triple() {
        assert_one("librs is a robust parser", "librs", "is-a", "parser");
        assert_one(
            "tantivy is an embedded library",
            "tantivy",
            "is-a",
            "library",
        );
        assert_one("sled is a concurrent database", "sled", "is-a", "database");
        // A modifier WordNet DOES give a noun sense now behaves identically,
        // which is the point: the lexicographic accident no longer changes the
        // outcome, only which token the re-walk skips.
        assert_one("librs is a fast parser", "librs", "is-a", "parser");
    }

    /// The lexicographic split the re-walk removed, measured rather than
    /// asserted by feel.
    ///
    /// Why: 43 of these 78 ordinary technical adjectives have no noun sense in
    /// WordNet 3.1. That is a property of Princeton's lexicography, not of
    /// anything tunable here, and it is the size of the blast radius the reject
    /// rule had. Pinned so a WordNet upgrade that moves it is visible.
    #[test]
    fn measures_the_adjective_only_population() {
        let wn = WordNetPos::shipped();
        let sample: &[&str] = &[
            "fast",
            "small",
            "simple",
            "modern",
            "lightweight",
            "portable",
            "generic",
            "old",
            "good",
            "great",
            "better",
            "new",
            "robust",
            "minimal",
            "tiny",
            "huge",
            "concurrent",
            "embedded",
            "reliable",
            "lazy",
            "secure",
            "rusty",
            "async",
            "asynchronous",
            "synchronous",
            "distributed",
            "persistent",
            "immutable",
            "mutable",
            "functional",
            "relational",
            "hierarchical",
            "incremental",
            "deterministic",
            "idempotent",
            "scalable",
            "extensible",
            "pluggable",
            "configurable",
            "optional",
            "required",
            "deprecated",
            "experimental",
            "stable",
            "unstable",
            "legacy",
            "native",
            "remote",
            "local",
            "static",
            "dynamic",
            "public",
            "private",
            "internal",
            "external",
            "open",
            "closed",
            "free",
            "paid",
            "commercial",
            "fancy",
            "neat",
            "clean",
            "dirty",
            "quick",
            "slow",
            "heavy",
            "light",
            "cheap",
            "expensive",
            "strict",
            "lenient",
            "safe",
            "unsafe",
            "correct",
            "incorrect",
            "complete",
            "partial",
        ];
        let adj_only = sample.iter().filter(|w| wn.is_adjective_only(w)).count();
        assert_eq!(
            (sample.len(), adj_only),
            (78, 43),
            "adjective-only population moved; re-measure before trusting the #5399 numbers"
        );
    }

    /// A phrase with no head names nothing, so no triple — and this is the
    /// re-walk returning `None`, not a surviving reject rule.
    #[test]
    fn an_all_adjective_subject_yields_no_triple() {
        assert_none("anything robust is a compiler");
    }

    /// The subject side gets the same re-walk. Without it, the old `last_token`
    /// would emit whichever token sat against the marker.
    #[test]
    fn subject_side_rewalks_past_an_adjective_only_token() {
        assert_one("the fast parser is a tool", "parser", "is-a", "tool");
    }

    // ====================== markdown and phrase boundaries ==================

    /// #5399: the terminator check used to read only the raw token's LAST
    /// character. `**MCP is a thin proxy.**` ends in `*`, so the `.` was missed
    /// and the run walked on into the next sentence, yielding
    /// `mcp --is-a--> sessions`. Markdown emphasis is the ordinary shape of
    /// drawer content, so this was never an edge case.
    #[test]
    fn stops_the_run_at_a_terminator_behind_markdown_emphasis() {
        assert_one(
            "**MCP is a thin proxy.** Sessions are cheap here",
            "mcp",
            "is-a",
            "proxy",
        );
    }

    /// The trailing-punctuation scan must not fire on interior punctuation, or
    /// every dotted name would close its own phrase.
    #[test]
    fn interior_punctuation_does_not_terminate_the_run() {
        for raw in ["parser", "node.js", "src/main.rs", "**bold**", "v1.2"] {
            assert!(!ends_noun_phrase(raw), "{raw:?} must not terminate");
        }
        for raw in ["proxy.**", "compiler,", "thing.", "group)", "list;", "end?"] {
            assert!(ends_noun_phrase(raw), "{raw:?} must terminate");
        }
    }

    /// Walking LEFT, a trailing `.` ended the PREVIOUS sentence, so the token
    /// carrying it belongs to that sentence and cannot join this phrase.
    /// Without the direction rule the head jumps the boundary and asserts
    /// `tantivy --is-a--> compiler` from two unrelated clauses.
    #[test]
    fn subject_walk_stops_before_a_previous_sentence() {
        assert_none("we shipped tantivy. robust is a compiler");
    }

    // ================= multi-line bodies, the production shape ==============
    // 🔴 Every other fixture in this module is a SINGLE LINE, and so is the
    // corpus harness (`examples/kg_dump.rs` reads one line per extraction).
    // Production is not: `tools::helpers::auto_extract_and_assert` and
    // `commands::kg_rebuild::rebuild_one` both pass a whole drawer body. The
    // walk had no newline boundary, so it ran off the end of its own sentence
    // and took a head from the next line — a correct-to-wrong regression that
    // no single-line fixture could see. Use `--whole-file` when measuring.

    /// A bare newline closes the phrase exactly as a period does.
    ///
    /// Against `8402bd8b` every row here produced the token from line two:
    /// `builds`, `builds`, `runs`, `sled`.
    #[test]
    fn the_object_walk_stops_at_a_line_break() {
        assert_one(
            "trusty-search is a daemon\ncargo builds it",
            "trusty-search",
            "is-a",
            "daemon",
        );
        assert_one(
            "rustc is a compiler\ncargo builds it",
            "rustc",
            "is-a",
            "compiler",
        );
        assert_one(
            "the parser is a tool\ncargo runs fine",
            "parser",
            "is-a",
            "tool",
        );
        assert_one(
            "redb is a database\nsled is another one",
            "redb",
            "is-a",
            "database",
        );
    }

    /// The subject side takes the same boundary: a preceding line is a
    /// previous sentence, whether or not it ended in punctuation.
    #[test]
    fn the_subject_walk_stops_at_a_line_break() {
        assert_one(
            "we shipped tantivy\nrobust compilers are a myth\nredb is a database",
            "redb",
            "is-a",
            "database",
        );
    }

    /// A trailing period already worked; it must keep working.
    #[test]
    fn a_terminated_line_is_unaffected_by_the_newline_rule() {
        assert_one(
            "trusty-search is a daemon.\ncargo builds it",
            "trusty-search",
            "is-a",
            "daemon",
        );
    }

    // ============ an inflected or unknown token must not take the head ======

    /// A participle is not a noun, and WordNet indexes only its base form.
    ///
    /// Against `8402bd8b` `containing` was unknown, and "unknown" meant
    /// "eligible to head a phrase", so this asserted
    /// `skill --is-a--> containing`. Resolving `containing` to `contain`
    /// (VERB-only) ends the phrase before it, as it always should have.
    #[test]
    fn a_participle_does_not_head_the_phrase() {
        assert_one(
            "each skill is a directory containing:",
            "skill",
            "is-a",
            "directory",
        );
        assert_one(
            "trusty-search is a daemon parsing every file",
            "trusty-search",
            "is-a",
            "daemon",
        );
    }

    /// An unknown token may OPEN a phrase but never joins one that already has
    /// a head — against `8402bd8b` this asserted the whole file path as a type.
    ///
    /// The closed-class `inside` is what stops the run here: it is a
    /// preposition, so the phrase ends before the path regardless of whether
    /// WordNet has heard of the path.
    #[test]
    fn an_unknown_token_does_not_displace_an_established_head() {
        assert_one(
            "tree is a comment inside crates/trusty-search/src/allowlist/tests.rs",
            "tree",
            "is-a",
            "comment",
        );
    }

    /// A plural subject survives, including one whose singular ends in `e`.
    ///
    /// Both of these yielded a triple at `8402bd8b` and nothing at `ba579925`:
    /// `base_form_candidates` chopped `es` before `s`, so `notes` resolved to
    /// the adverb `not` and `sites` to the verb `sit`, and a verb-or-adverb
    /// token ends the phrase. The sibilant rule in
    /// [`crate::wordnet_pos::base_form_candidates`] is what restores them.
    #[test]
    fn a_plural_subject_whose_singular_ends_in_e_still_heads_its_phrase() {
        assert_one("notes is a drawer", "notes", "is-a", "drawer");
        assert_one("sites is a directory", "sites", "is-a", "directory");
    }

    /// An unknown identifier still heads its phrase when the token before it is
    /// a modifier WordNet also lists as a noun.
    ///
    /// 🔴 DO NOT "fix" this by stopping the walk at an unknown token once the
    /// run is non-empty. That rule was written and then removed here, measured
    /// over 1,277 markdown files: it changed 156 heads, and 83 of those
    /// replaced the real head with an adjective-capable modifier —
    /// `tcode|is-a|local`, `persistence|is-a|redb` -> `single`,
    /// `file|is-a|no-op` -> `true`. That is the #5399 defect itself arriving by
    /// a different route, and [`phrase_head`]'s adjective-only skip cannot
    /// catch it, because WordNet lists every one of those words as a noun as
    /// well. It also dropped 76 triples outright, several of them real facts
    /// (`server|uses|rocksdb`, `codebase|uses|redb`).
    /// Test: this is the isolating test — reintroducing the rule fails it.
    #[test]
    fn an_unknown_token_heads_its_phrase_over_an_adjective_capable_modifier() {
        // `local` is NOUN|ADJ, so it passes both the POS check and the
        // adjective-only skip; only walking on to `app` finds the real head.
        assert_one("tcode is a local app", "tcode", "is-a", "app");
        assert_one(
            "persistence is a single redb",
            "persistence",
            "is-a",
            "redb",
        );
    }

    #[test]
    fn stops_the_noun_phrase_run_at_a_comma() {
        assert_one(
            "rustc is a compiler, and cargo is the build tool",
            "rustc",
            "is-a",
            "compiler",
        );
    }

    #[test]
    fn caps_the_noun_phrase_run() {
        // Five noun/adjective tokens; the head must come from the first four.
        assert_one(
            "librs is a fast small modular parser core",
            "librs",
            "is-a",
            "parser",
        );
    }

    // ============================ fail-open =================================

    #[test]
    fn unknown_subject_and_object_both_fail_open() {
        assert_one("tantivy uses fst", "tantivy", "uses", "fst");
    }

    /// WordNet knows ordinary English, and ordinary English words are also
    /// language and tool names. Nothing here rejects a known word, so these
    /// work — but the fail-open rule is what saves them, and any future
    /// "reject known common nouns" tightening would break exactly these.
    #[test]
    fn common_english_words_that_are_real_entity_names_survive() {
        assert_one("rust is a language", "rust", "is-a", "language");
        assert_one("go is a language", "go", "is-a", "language");
        assert_one("python is a language", "python", "is-a", "language");
    }

    /// The caller's table is the one consulted — the property that replaced the
    /// `OnceLock`. Against a table that knows nothing, every token fails open,
    /// so the walk stops at the first one and reproduces the #4678 residue on
    /// demand. That is the proof the shipped table is what removes it.
    #[test]
    fn a_caller_supplied_pos_table_is_the_one_consulted() {
        let config = KgExtractConfig {
            pos: WordNetPos::from_table("zzz\t1\n"),
            ..Default::default()
        };
        let got: Vec<_> = extract_triples_with_config(
            &ExtractInput {
                drawer_id: Uuid::new_v4(),
                content: "match exhaustiveness is a hard requirement here",
                tags: &[],
                room: None,
            },
            &config,
        )
        .into_iter()
        .filter(|t| PATTERN_TABLE.iter().any(|(p, _)| *p == t.predicate))
        .map(|t| (t.subject, t.predicate, t.object))
        .collect();
        assert_eq!(
            got,
            vec![(
                "exhaustiveness".to_string(),
                "is-a".to_string(),
                "hard".to_string()
            )]
        );
    }

    // ================= where this approach behaves surprisingly =============
    // Each of these is a KNOWN limitation, filed separately and deliberately
    // NOT fixed in #5399. They are pinned so that a future change altering them
    // does so on purpose.

    /// Only `of` is treated as NP-continuing, so the same truncation slips
    /// through with every other preposition. Widening the set breaks row 5
    /// (`uses redb for persistence`); what actually separates the two is
    /// relational-noun complement requirements, which WordNet does not encode.
    #[test]
    fn surprising_non_genitive_relational_phrase_still_truncates() {
        assert_one(
            "the daemon is a participant in the process group",
            "daemon",
            "is-a",
            "participant",
        );
    }

    /// A regular inflection is POS-checked through its base form.
    ///
    /// 🔴 THIS EXPECTATION WAS CHANGED, AND THE OLD ONE DESCRIBED A DEFECT.
    /// It used to assert `mask("parsers") == 0` and file that under "known
    /// limitations, benign" — WordNet indexes base forms, so an inflected word
    /// simply took the fail-open path. That reading missed what fail-open
    /// MEANS inside the walk: unknown is what makes a token eligible to head a
    /// phrase, so every participle became a candidate head and won the slot by
    /// sitting rightmost. `a directory containing:` asserted `containing` as
    /// the type. So the limitation was not benign, and the fix is to resolve
    /// the base form rather than to refuse unknown tokens — refusing them
    /// outright would take `parsers` with it, and `parsers` is a real head.
    #[test]
    fn inflected_forms_resolve_through_their_base_form() {
        let wn = WordNetPos::shipped();
        assert!(
            wn.is_noun("parsers"),
            "a regular plural keeps its noun sense"
        );
        assert_eq!(wn.mask("containing") & wordnet_pos::NOUN, 0);
        assert_one("librs is a fast parsers", "librs", "is-a", "parsers");
    }

    /// A gerund that WordNet lists as a noun in its own right still takes the
    /// head, because membership cannot tell a nominalisation from a participle.
    ///
    /// `running` is NOUN|ADJ and `clearing` is NOUN, so no base-form retry
    /// fires and both read as ordinary nouns. Stopping the run at any `-ing`
    /// word whose base is a verb WOULD fix these two, and it was measured over
    /// this repo's markdown before being rejected. The count first recorded
    /// here — "repairs ~25, breaks ~18" — was wrong and flattered the decision;
    /// the break side is the larger one. Over the 1,277-file corpus the rule
    /// reaches 71 head positions ending in `-ing` (47 distinct), and most of
    /// them are ordinary nouns it would destroy: `running`, `mapping`,
    /// `warning`, `tooling`, `ranking`, `understanding`. Telling a
    /// nominalisation from a participle is syntax, not lexical membership, so
    /// it needs a tagger rather than a wider suffix table.
    #[test]
    fn surprising_a_gerund_noun_still_takes_the_head() {
        assert_one(
            "session creation depends on daemon running",
            "creation",
            "depends-on",
            "running",
        );
        assert_one(
            "budget_tokens is an estimate clearing a budget",
            "budget_tokens",
            "is-a",
            "clearing",
        );
    }

    /// Two unknown tokens in a row: the run stops at the first, so the head is
    /// the first name rather than the last. Right for `uses redb for X`, wrong
    /// for a genuine unknown-unknown compound.
    #[test]
    fn surprising_run_stops_at_the_first_unknown_token() {
        assert_one(
            "trusty-memory uses redb sled",
            "trusty-memory",
            "uses",
            "redb",
        );
    }

    /// A single-token head truncates a compound.
    #[test]
    fn surprising_single_token_head_truncates_a_compound() {
        assert_one(
            "brew is a system package managers",
            "brew",
            "is-a",
            "managers",
        );
    }
}