Skip to main content

panproto_parse/
emit_pretty.rs

1#![allow(
2    clippy::module_name_repetitions,
3    clippy::too_many_lines,
4    clippy::too_many_arguments,
5    clippy::map_unwrap_or,
6    clippy::option_if_let_else,
7    clippy::elidable_lifetime_names,
8    clippy::items_after_statements,
9    clippy::needless_pass_by_value,
10    clippy::single_match_else,
11    clippy::manual_let_else,
12    clippy::match_same_arms,
13    clippy::missing_const_for_fn,
14    clippy::single_char_pattern,
15    clippy::naive_bytecount,
16    clippy::expect_used,
17    clippy::redundant_pub_crate,
18    clippy::used_underscore_binding,
19    clippy::redundant_field_names,
20    clippy::struct_field_names,
21    clippy::redundant_else,
22    clippy::similar_names
23)]
24
25//! De-novo source emission from a by-construction schema.
26//!
27//! [`AstParser::emit`] reconstructs source from byte-position fragments
28//! that the parser stored on the schema during `parse`. That works for
29//! edit pipelines (`parse → transform → emit`) but fails for schemas
30//! built by hand (`SchemaBuilder` with no parse history): they carry
31//! no `start-byte`, no `interstitial-N`, no `literal-value`, and the
32//! reconstructor returns `Err(EmitFailed { reason: "schema has no
33//! text fragments" })`.
34//!
35//! This module renders such schemas to source bytes by walking
36//! tree-sitter's `grammar.json` production rules. For each schema
37//! vertex of kind `K`, the walker looks up `K`'s production in the
38//! grammar and emits its body in order:
39//!
40//! - `STRING` nodes contribute literal token bytes directly.
41//! - `SYMBOL` and `FIELD` nodes recurse into the schema's children,
42//!   matching by edge kind (which is the tree-sitter field name).
43//! - `SEQ` emits its members in order.
44//! - `CHOICE` picks the alternative whose head `SYMBOL` matches an
45//!   actual child kind, or whose terminals appear in the rendered
46//!   prefix; falls back to the first non-`BLANK` alternative when no
47//!   alternative matches.
48//! - `REPEAT` and `REPEAT1` emit their content once per matching
49//!   child edge in declared order.
50//! - `OPTIONAL` emits its content iff a corresponding child edge or
51//!   constraint is populated.
52//! - `PATTERN` is a regex placeholder for variable-text terminals
53//!   (identifiers, numbers, quoted strings). The walker emits a
54//!   `literal-value` constraint when present and otherwise falls
55//!   back to a placeholder derived from the regex shape.
56//! - `BLANK`, `TOKEN`, `IMMEDIATE_TOKEN`, `ALIAS`, `PREC*` are
57//!   handled transparently (the inner content is emitted; the
58//!   wrapper is dropped).
59//!
60//! Whitespace and indentation come from a `FormatPolicy` applied
61//! during emission. The default policy inserts a single space between
62//! adjacent tokens, a newline after `;` / `}` / `{`, and tracks an
63//! indent counter on `{` / `}` boundaries.
64//!
65//! Output is *syntactically valid* for any grammar that ships
66//! `grammar.json`. Idiomatic formatting (rustfmt-style spacing rules,
67//! per-language conventions) is a polish layer that lives outside
68//! this module.
69
70use std::collections::BTreeMap;
71
72use panproto_schema::{Edge, Schema};
73use serde::Deserialize;
74
75use crate::error::ParseError;
76
77// ═══════════════════════════════════════════════════════════════════
78// Grammar JSON model
79// ═══════════════════════════════════════════════════════════════════
80
81/// A single tree-sitter production rule.
82///
83/// Mirrors the shape emitted by `tree-sitter generate`: every node has
84/// a `type` discriminator that selects a structural variant. The
85/// untyped subset (`PATTERN`, `STRING`, `SYMBOL`, `BLANK`) handles
86/// terminals; the structural subset (`SEQ`, `CHOICE`, `REPEAT`,
87/// `REPEAT1`, `OPTIONAL`, `FIELD`, `ALIAS`, `TOKEN`,
88/// `IMMEDIATE_TOKEN`, `PREC*`) builds composite productions.
89#[derive(Debug, Clone, Deserialize)]
90#[serde(tag = "type")]
91#[non_exhaustive]
92pub enum Production {
93    /// Concatenation of productions.
94    #[serde(rename = "SEQ")]
95    Seq {
96        /// Ordered members; each is emitted in turn.
97        members: Vec<Self>,
98    },
99    /// Alternation between productions.
100    #[serde(rename = "CHOICE")]
101    Choice {
102        /// Alternatives; the walker picks one based on the schema's
103        /// children and constraints.
104        members: Vec<Self>,
105    },
106    /// Zero-or-more repetition.
107    #[serde(rename = "REPEAT")]
108    Repeat {
109        /// The repeated body.
110        content: Box<Self>,
111    },
112    /// One-or-more repetition.
113    #[serde(rename = "REPEAT1")]
114    Repeat1 {
115        /// The repeated body.
116        content: Box<Self>,
117    },
118    /// Optional inclusion (zero or one).
119    ///
120    /// Tree-sitter usually emits `OPTIONAL` as `CHOICE { content,
121    /// BLANK }`, but recent generator versions also emit explicit
122    /// `OPTIONAL` nodes; both shapes are accepted.
123    #[serde(rename = "OPTIONAL")]
124    Optional {
125        /// The optional body.
126        content: Box<Self>,
127    },
128    /// Reference to another rule by name.
129    #[serde(rename = "SYMBOL")]
130    Symbol {
131        /// Name of the referenced rule (matches a vertex kind on the
132        /// schema side).
133        name: String,
134    },
135    /// Literal token bytes.
136    #[serde(rename = "STRING")]
137    String {
138        /// The literal token. Emitted verbatim.
139        value: String,
140    },
141    /// Regex-matched terminal.
142    ///
143    /// At parse time this matches arbitrary bytes; at emit time the
144    /// walker substitutes a `literal-value` constraint when present
145    /// and falls back to a placeholder otherwise.
146    #[serde(rename = "PATTERN")]
147    Pattern {
148        /// The original regex.
149        value: String,
150    },
151    /// The empty production. Emits nothing.
152    #[serde(rename = "BLANK")]
153    Blank,
154    /// Named field over a content production.
155    ///
156    /// The field `name` matches an edge kind on the schema side; the
157    /// walker resolves the corresponding child vertex and recurses
158    /// into `content` with that child as context.
159    #[serde(rename = "FIELD")]
160    Field {
161        /// Field name (matches edge kind).
162        name: String,
163        /// The contents of the field.
164        content: Box<Self>,
165    },
166    /// An aliased production.
167    ///
168    /// `value` records the parser-visible kind; the walker emits
169    /// `content` and ignores the alias rename.
170    #[serde(rename = "ALIAS")]
171    Alias {
172        /// The aliased content.
173        content: Box<Self>,
174        /// Whether the alias is a named node.
175        #[serde(default)]
176        named: bool,
177        /// The alias's surface name.
178        #[serde(default)]
179        value: String,
180    },
181    /// A token wrapper.
182    ///
183    /// Tree-sitter uses `TOKEN` to mark a sub-rule as a single
184    /// lexical token; the walker emits the inner content unchanged.
185    #[serde(rename = "TOKEN")]
186    Token {
187        /// The wrapped content.
188        content: Box<Self>,
189    },
190    /// An immediate-token wrapper (no preceding whitespace).
191    ///
192    /// Treated like [`Production::Token`] for emit purposes.
193    #[serde(rename = "IMMEDIATE_TOKEN")]
194    ImmediateToken {
195        /// The wrapped content.
196        content: Box<Self>,
197    },
198    /// Precedence wrapper.
199    #[serde(rename = "PREC")]
200    Prec {
201        /// Precedence value (numeric or string). Ignored at emit time.
202        #[allow(dead_code)]
203        value: serde_json::Value,
204        /// The wrapped content.
205        content: Box<Self>,
206    },
207    /// Left-associative precedence wrapper.
208    #[serde(rename = "PREC_LEFT")]
209    PrecLeft {
210        /// Precedence value. Ignored at emit time.
211        #[allow(dead_code)]
212        value: serde_json::Value,
213        /// The wrapped content.
214        content: Box<Self>,
215    },
216    /// Right-associative precedence wrapper.
217    #[serde(rename = "PREC_RIGHT")]
218    PrecRight {
219        /// Precedence value. Ignored at emit time.
220        #[allow(dead_code)]
221        value: serde_json::Value,
222        /// The wrapped content.
223        content: Box<Self>,
224    },
225    /// Dynamic precedence wrapper.
226    #[serde(rename = "PREC_DYNAMIC")]
227    PrecDynamic {
228        /// Precedence value. Ignored at emit time.
229        #[allow(dead_code)]
230        value: serde_json::Value,
231        /// The wrapped content.
232        content: Box<Self>,
233    },
234    /// Reserved-word wrapper (tree-sitter ≥ 0.25).
235    ///
236    /// Tree-sitter's `RESERVED` rule marks an inner production as a
237    /// reserved-word context: the parser excludes the listed identifiers
238    /// from being treated as the inner symbol. The `context_name`
239    /// metadata names the reserved-word set; the emitter does not need
240    /// it (we are walking schema → bytes, not enforcing reserved-word
241    /// constraints), so we emit the inner content unchanged, the same
242    /// way [`Production::Token`] and [`Production::ImmediateToken`] do.
243    #[serde(rename = "RESERVED")]
244    Reserved {
245        /// The wrapped content.
246        content: Box<Self>,
247        /// Name of the reserved-word context. Ignored at emit time.
248        #[allow(dead_code)]
249        #[serde(default)]
250        context_name: String,
251    },
252}
253
254/// A grammar's production-rule table, deserialized from `grammar.json`.
255///
256/// Only the fields the emitter consumes are decoded; precedences,
257/// conflicts, externals, and other parser-only metadata are ignored.
258#[derive(Debug, Clone, Deserialize)]
259pub struct Grammar {
260    /// Grammar name (e.g. `"rust"`, `"typescript"`).
261    #[allow(dead_code)]
262    pub name: String,
263    /// Map from rule name (a vertex kind on the schema side) to
264    /// production. Entries are kept in lexical order so iteration
265    /// is deterministic.
266    pub rules: BTreeMap<String, Production>,
267    /// Supertypes declared in the grammar's `supertypes` field. A
268    /// supertype is a rule whose body is a `CHOICE` of `SYMBOL`
269    /// references; tree-sitter parsers report a node's kind as one
270    /// of the subtypes (e.g. `identifier`, `typed_parameter`) rather
271    /// than the supertype name (`parameter`), so the emitter needs to
272    /// know that a child kind in a subtype set should match the
273    /// supertype name when a SYMBOL references it.
274    #[serde(default, deserialize_with = "deserialize_supertypes")]
275    pub supertypes: std::collections::HashSet<String>,
276    /// Precomputed subtyping closure: `subtypes[symbol_name]` is the
277    /// set of vertex kinds that satisfy a SYMBOL `symbol_name`
278    /// reference on the schema side.
279    ///
280    /// Built once at [`Grammar::from_bytes`] time by walking each
281    /// hidden rule (`_`-prefixed), declared supertype, and named
282    /// `ALIAS { value: K, ... }` production to its leaf SYMBOLs and
283    /// recording the closure. This replaces the prior heuristic
284    /// `kind_satisfies_symbol` that walked the rule body on every
285    /// query: lookups are now O(1) and the relation is exactly the
286    /// transitive closure of "is reachable via hidden / supertype /
287    /// alias dispatch", with no over-expansion through non-hidden
288    /// non-supertype rule references.
289    #[serde(skip)]
290    pub subtypes: std::collections::HashMap<String, std::collections::HashSet<String>>,
291}
292
293fn deserialize_supertypes<'de, D>(
294    deserializer: D,
295) -> Result<std::collections::HashSet<String>, D::Error>
296where
297    D: serde::Deserializer<'de>,
298{
299    let entries: Vec<serde_json::Value> = Vec::deserialize(deserializer)?;
300    let mut out = std::collections::HashSet::new();
301    for entry in entries {
302        match entry {
303            serde_json::Value::String(s) => {
304                out.insert(s);
305            }
306            serde_json::Value::Object(map) => {
307                if let Some(serde_json::Value::String(name)) = map.get("name") {
308                    out.insert(name.clone());
309                }
310            }
311            _ => {}
312        }
313    }
314    Ok(out)
315}
316
317impl Grammar {
318    /// Parse a grammar's `grammar.json` bytes.
319    ///
320    /// Builds the subtyping closure as part of construction so every
321    /// downstream lookup is O(1). The closure is the least relation
322    /// containing `(K, K)` for every rule key `K` and closed under:
323    ///
324    /// - hidden-rule expansion: if `S` is hidden and a SYMBOL `S` may
325    ///   reach SYMBOL `K`, then `K ⊑ S`.
326    /// - supertype expansion: if `S` is in the grammar's supertypes
327    ///   block and `K` is one of `S`'s alternatives, then `K ⊑ S`.
328    /// - alias renaming: if a rule body contains
329    ///   `ALIAS { content: SYMBOL R, value: A, named: true }` where
330    ///   `R` reaches kind `K` (or `K = R` when no further hop), then
331    ///   `A ⊑ R` and `K ⊑ A`.
332    ///
333    /// # Errors
334    ///
335    /// Returns [`ParseError::EmitFailed`] when the bytes are not a
336    /// valid `grammar.json` document.
337    pub fn from_bytes(protocol: &str, bytes: &[u8]) -> Result<Self, ParseError> {
338        let mut grammar: Self =
339            serde_json::from_slice(bytes).map_err(|e| ParseError::EmitFailed {
340                protocol: protocol.to_owned(),
341                reason: format!("grammar.json deserialization failed: {e}"),
342            })?;
343        grammar.subtypes = compute_subtype_closure(&grammar);
344        Ok(grammar)
345    }
346}
347
348/// Compute the subtyping relation as a forward-indexed map from a
349/// SYMBOL name to the set of vertex kinds that satisfy that SYMBOL.
350fn compute_subtype_closure(
351    grammar: &Grammar,
352) -> std::collections::HashMap<String, std::collections::HashSet<String>> {
353    use std::collections::{HashMap, HashSet};
354    // Edges of the "kind X satisfies SYMBOL Y" relation. `K ⊑ Y` is
355    // recorded whenever Y is reached by walking the grammar's
356    // ALIAS / hidden-rule / supertype dispatch from a position where
357    // K is the actual vertex kind.
358    let mut subtypes: HashMap<String, HashSet<String>> = HashMap::new();
359    for name in grammar.rules.keys() {
360        subtypes
361            .entry(name.clone())
362            .or_default()
363            .insert(name.clone());
364    }
365
366    // First pass: collect the immediate "satisfies" edges from each
367    // expandable rule (hidden, supertype) to the kinds reachable by
368    // walking its body, plus alias edges.
369    fn walk<'g>(
370        grammar: &'g Grammar,
371        production: &'g Production,
372        visited: &mut HashSet<&'g str>,
373        out: &mut HashSet<String>,
374    ) {
375        match production {
376            Production::Symbol { name } => {
377                // Direct subtype.
378                out.insert(name.clone());
379                // Continue expansion through hidden / supertype rules
380                // so the closure traverses pass-through dispatch.
381                let expand = name.starts_with('_') || grammar.supertypes.contains(name.as_str());
382                if expand && visited.insert(name.as_str()) {
383                    if let Some(rule) = grammar.rules.get(name) {
384                        walk(grammar, rule, visited, out);
385                    }
386                }
387            }
388            Production::Choice { members } | Production::Seq { members } => {
389                for m in members {
390                    walk(grammar, m, visited, out);
391                }
392            }
393            Production::Alias {
394                content,
395                named,
396                value,
397            } => {
398                if *named && !value.is_empty() {
399                    out.insert(value.clone());
400                }
401                walk(grammar, content, visited, out);
402            }
403            Production::Repeat { content }
404            | Production::Repeat1 { content }
405            | Production::Optional { content }
406            | Production::Field { content, .. }
407            | Production::Token { content }
408            | Production::ImmediateToken { content }
409            | Production::Prec { content, .. }
410            | Production::PrecLeft { content, .. }
411            | Production::PrecRight { content, .. }
412            | Production::PrecDynamic { content, .. }
413            | Production::Reserved { content, .. } => {
414                walk(grammar, content, visited, out);
415            }
416            _ => {}
417        }
418    }
419
420    for (name, rule) in &grammar.rules {
421        let expand = name.starts_with('_') || grammar.supertypes.contains(name.as_str());
422        if !expand {
423            continue;
424        }
425        let mut visited: HashSet<&str> = HashSet::new();
426        visited.insert(name.as_str());
427        let mut reachable: HashSet<String> = HashSet::new();
428        walk(grammar, rule, &mut visited, &mut reachable);
429        for kind in &reachable {
430            subtypes
431                .entry(kind.clone())
432                .or_default()
433                .insert(name.clone());
434        }
435    }
436
437    // Aliases: scan every rule body for ALIAS { content, value }
438    // declarations. The kinds reachable from `content` satisfy
439    // `value`, AND (by construction) `value` satisfies the
440    // surrounding rule. Walking the ENTIRE grammar once captures
441    // every alias site, irrespective of which rule introduces it.
442    fn collect_aliases<'g>(production: &'g Production, out: &mut Vec<(String, &'g Production)>) {
443        match production {
444            Production::Alias {
445                content,
446                named,
447                value,
448            } => {
449                if *named && !value.is_empty() {
450                    out.push((value.clone(), content.as_ref()));
451                }
452                collect_aliases(content, out);
453            }
454            Production::Choice { members } | Production::Seq { members } => {
455                for m in members {
456                    collect_aliases(m, out);
457                }
458            }
459            Production::Repeat { content }
460            | Production::Repeat1 { content }
461            | Production::Optional { content }
462            | Production::Field { content, .. }
463            | Production::Token { content }
464            | Production::ImmediateToken { content }
465            | Production::Prec { content, .. }
466            | Production::PrecLeft { content, .. }
467            | Production::PrecRight { content, .. }
468            | Production::PrecDynamic { content, .. }
469            | Production::Reserved { content, .. } => {
470                collect_aliases(content, out);
471            }
472            _ => {}
473        }
474    }
475    let mut aliases: Vec<(String, &Production)> = Vec::new();
476    for rule in grammar.rules.values() {
477        collect_aliases(rule, &mut aliases);
478    }
479    for (alias_value, content) in aliases {
480        let mut visited: HashSet<&str> = HashSet::new();
481        let mut reachable: HashSet<String> = HashSet::new();
482        walk(grammar, content, &mut visited, &mut reachable);
483        // Aliased value satisfies itself and is satisfied by every
484        // kind its content can reach.
485        subtypes
486            .entry(alias_value.clone())
487            .or_default()
488            .insert(alias_value.clone());
489        for kind in reachable {
490            subtypes
491                .entry(kind)
492                .or_default()
493                .insert(alias_value.clone());
494        }
495    }
496
497    // Transitive close: `K ⊑ A` and `A ⊑ B` implies `K ⊑ B`. Iterate
498    // a few rounds; the relation is small so a quick fixed-point
499    // suffices in practice.
500    for _ in 0..8 {
501        let snapshot = subtypes.clone();
502        let mut changed = false;
503        for (kind, supers) in &snapshot {
504            let extra: HashSet<String> = supers
505                .iter()
506                .flat_map(|s| snapshot.get(s).cloned().unwrap_or_default())
507                .collect();
508            let entry = subtypes.entry(kind.clone()).or_default();
509            for s in extra {
510                if entry.insert(s) {
511                    changed = true;
512                }
513            }
514        }
515        if !changed {
516            break;
517        }
518    }
519
520    subtypes
521}
522
523// ═══════════════════════════════════════════════════════════════════
524// Format policy
525// ═══════════════════════════════════════════════════════════════════
526
527/// Whitespace and indentation policy applied during emission.
528///
529/// The default policy inserts a single space between adjacent tokens,
530/// a newline after `;` / `}` / `{`, and tracks indent on `{` / `}`
531/// boundaries. Per-language overrides (idiomatic indent width,
532/// trailing-comma rules, blank-line conventions) can ride alongside
533/// this struct in a follow-up branch; today's defaults aim only for
534/// syntactic validity.
535#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
536pub struct FormatPolicy {
537    /// Number of spaces per indent level.
538    pub indent_width: usize,
539    /// Separator inserted between adjacent terminals that the lexer
540    /// would otherwise glue together (word ↔ word, operator ↔ operator).
541    /// Default is a single space.
542    pub separator: String,
543    /// Newline byte sequence emitted after `line_break_after` tokens
544    /// and at end-of-output. Default is `"\n"`.
545    pub newline: String,
546    /// Tokens after which the walker breaks to a new line.
547    pub line_break_after: Vec<String>,
548    /// Tokens that increase indent on emission.
549    pub indent_open: Vec<String>,
550    /// Tokens that decrease indent on emission.
551    pub indent_close: Vec<String>,
552}
553
554impl Default for FormatPolicy {
555    fn default() -> Self {
556        Self {
557            indent_width: 2,
558            separator: " ".to_owned(),
559            newline: "\n".to_owned(),
560            line_break_after: vec![";".into(), "{".into(), "}".into()],
561            indent_open: vec!["{".into()],
562            indent_close: vec!["}".into()],
563        }
564    }
565}
566
567// ═══════════════════════════════════════════════════════════════════
568// Emitter
569// ═══════════════════════════════════════════════════════════════════
570
571/// Emit a by-construction schema to source bytes.
572///
573/// `protocol` is the grammar / language name (used in error messages
574/// and to label the entry point).
575///
576/// The walker treats `schema.entries` as the ordered list of root
577/// vertices, falling back to a deterministic by-id ordering when
578/// `entries` is empty. Each root is emitted using the production
579/// associated with its kind in `grammar.rules`.
580///
581/// # Errors
582///
583/// Returns [`ParseError::EmitFailed`] when:
584///
585/// - the schema has no vertices
586/// - a root vertex's kind is not a grammar rule
587/// - a `SYMBOL` reference points at a kind with no rule and no schema
588///   child to resolve it to
589/// - a required `FIELD` has no corresponding edge in the schema
590pub fn emit_pretty(
591    protocol: &str,
592    schema: &Schema,
593    grammar: &Grammar,
594    policy: &FormatPolicy,
595) -> Result<Vec<u8>, ParseError> {
596    let roots = collect_roots(schema);
597    if roots.is_empty() {
598        return Err(ParseError::EmitFailed {
599            protocol: protocol.to_owned(),
600            reason: "schema has no entry vertices".to_owned(),
601        });
602    }
603
604    let mut out = Output::new(policy);
605    for (i, root) in roots.iter().enumerate() {
606        if i > 0 {
607            out.newline();
608        }
609        emit_vertex(protocol, schema, grammar, root, &mut out)?;
610    }
611    Ok(out.finish())
612}
613
614fn collect_roots(schema: &Schema) -> Vec<&panproto_gat::Name> {
615    if !schema.entries.is_empty() {
616        return schema
617            .entries
618            .iter()
619            .filter(|name| schema.vertices.contains_key(*name))
620            .collect();
621    }
622
623    // Fallback: every vertex that is not the target of any structural edge
624    // (sorted by id for determinism).
625    let mut targets: std::collections::HashSet<&panproto_gat::Name> =
626        std::collections::HashSet::new();
627    for edge in schema.edges.keys() {
628        targets.insert(&edge.tgt);
629    }
630    let mut roots: Vec<&panproto_gat::Name> = schema
631        .vertices
632        .keys()
633        .filter(|name| !targets.contains(name))
634        .collect();
635    roots.sort();
636    roots
637}
638
639fn emit_vertex(
640    protocol: &str,
641    schema: &Schema,
642    grammar: &Grammar,
643    vertex_id: &panproto_gat::Name,
644    out: &mut Output<'_>,
645) -> Result<(), ParseError> {
646    let vertex = schema
647        .vertices
648        .get(vertex_id)
649        .ok_or_else(|| ParseError::EmitFailed {
650            protocol: protocol.to_owned(),
651            reason: format!("vertex '{vertex_id}' not found"),
652        })?;
653
654    // Leaf shortcut: a vertex carrying a `literal-value` constraint
655    // and no outgoing structural edges is a terminal token. Emit the
656    // captured value directly. This handles identifiers, numeric
657    // literals, and string literals that the parser stored as
658    // `literal-value` even on by-construction schemas.
659    if let Some(literal) = literal_value(schema, vertex_id) {
660        if children_for(schema, vertex_id).is_empty() {
661            out.token(literal);
662            return Ok(());
663        }
664    }
665
666    let kind = vertex.kind.as_ref();
667    let edges = children_for(schema, vertex_id);
668    if let Some(rule) = grammar.rules.get(kind) {
669        let mut cursor = ChildCursor::new(&edges);
670        return emit_production(protocol, schema, grammar, vertex_id, rule, &mut cursor, out);
671    }
672
673    // No rule for this kind. The parser produced it via an ALIAS
674    // (tree-sitter's `alias($.something, $.actual_kind)`) or via an
675    // external scanner (e.g. YAML's `document` root). Fall back to
676    // walking the children directly so the inner content survives;
677    // surrounding tokens — whose only source is the missing rule —
678    // are necessarily absent.
679    for edge in &edges {
680        emit_vertex(protocol, schema, grammar, &edge.tgt, out)?;
681    }
682    Ok(())
683}
684
685/// Linear cursor over a vertex's outgoing edges, used to thread
686/// children through a production rule without double-consuming them.
687struct ChildCursor<'a> {
688    edges: &'a [&'a Edge],
689    consumed: Vec<bool>,
690}
691
692impl<'a> ChildCursor<'a> {
693    fn new(edges: &'a [&'a Edge]) -> Self {
694        Self {
695            edges,
696            consumed: vec![false; edges.len()],
697        }
698    }
699
700    /// Take the next unconsumed edge whose kind equals `field_name`.
701    fn take_field(&mut self, field_name: &str) -> Option<&'a Edge> {
702        for (i, edge) in self.edges.iter().enumerate() {
703            if !self.consumed[i] && edge.kind.as_ref() == field_name {
704                self.consumed[i] = true;
705                return Some(edge);
706            }
707        }
708        None
709    }
710
711    /// Whether any unconsumed edge satisfies `predicate`. Used by the
712    /// unit tests; the live emit path went through `has_matching` on
713    /// each alternative until cursor-driven dispatch was rewritten to
714    /// pick the first-unconsumed-edge's kind directly.
715    #[cfg(test)]
716    fn has_matching(&self, predicate: impl Fn(&Edge) -> bool) -> bool {
717        self.edges
718            .iter()
719            .enumerate()
720            .any(|(i, edge)| !self.consumed[i] && predicate(edge))
721    }
722
723    /// Take the next unconsumed edge whose target vertex satisfies
724    /// `predicate`. Returns the edge and the underlying production
725    /// resolution path is the caller's job.
726    fn take_matching(&mut self, predicate: impl Fn(&Edge) -> bool) -> Option<&'a Edge> {
727        for (i, edge) in self.edges.iter().enumerate() {
728            if !self.consumed[i] && predicate(edge) {
729                self.consumed[i] = true;
730                return Some(edge);
731            }
732        }
733        None
734    }
735}
736
737thread_local! {
738    static EMIT_DEPTH: std::cell::Cell<usize> = const { std::cell::Cell::new(0) };
739    /// Set of `(vertex_id, rule_name)` pairs that are currently being
740    /// walked by the recursion. A SYMBOL that resolves to a rule
741    /// already on this stack closes a μ-binder cycle: in the
742    /// coinductive reading, the rule walk at any vertex is the least
743    /// fixed point of `body[μ X . body / X]`, which unfolds at most
744    /// once, with the second visit returning the empty sequence (the
745    /// unit of the free token monoid). Examples that trigger this:
746    /// YAML's `stream` ⊃ `_b_blk_*` mutually-recursive chain, Rust's
747    /// `_expression` ⊃ `binary_expression` ⊃ `_expression`.
748    static EMIT_MU_FRAMES: std::cell::RefCell<std::collections::HashSet<(String, String)>> =
749        std::cell::RefCell::new(std::collections::HashSet::new());
750    /// The name of the FIELD whose body the walker is currently inside,
751    /// or `None` at top level. Lets a SYMBOL nested arbitrarily deep
752    /// in the field's content (under SEQ, CHOICE, REPEAT, OPTIONAL)
753    /// consume from the *outer* cursor by edge-kind rather than from
754    /// the child's own cursor by symbol-match. Without this, shapes
755    /// like `field('args', commaSep1($.X))` — which expands to
756    /// `FIELD(SEQ(SYMBOL X, REPEAT(SEQ(',', SYMBOL X))))` — emit only
757    /// the first matched edge: the FIELD handler consumed one edge,
758    /// the inner REPEAT searched the consumed child's cursor (which
759    /// has no more sibling field edges), and the REPEAT broke after
760    /// one iteration. Setting the context here so the inner SYMBOL
761    /// pulls successive field-named edges from the outer cursor
762    /// recovers every matched edge across arbitrary nesting.
763    static EMIT_FIELD_CONTEXT: std::cell::RefCell<Option<String>> =
764        const { std::cell::RefCell::new(None) };
765}
766
767/// RAII guard that restores the prior `EMIT_FIELD_CONTEXT` value on drop.
768struct FieldContextGuard(Option<String>);
769
770impl Drop for FieldContextGuard {
771    fn drop(&mut self) {
772        EMIT_FIELD_CONTEXT.with(|f| *f.borrow_mut() = self.0.take());
773    }
774}
775
776fn push_field_context(name: &str) -> FieldContextGuard {
777    let prev = EMIT_FIELD_CONTEXT.with(|f| f.borrow_mut().replace(name.to_owned()));
778    FieldContextGuard(prev)
779}
780
781/// Clear the field context for the duration of a child-context walk.
782/// The child's own production has its own FIELDs that set their own
783/// context; the outer field hint must not leak into them.
784fn clear_field_context() -> FieldContextGuard {
785    let prev = EMIT_FIELD_CONTEXT.with(|f| f.borrow_mut().take());
786    FieldContextGuard(prev)
787}
788
789fn current_field_context() -> Option<String> {
790    EMIT_FIELD_CONTEXT.with(|f| f.borrow().clone())
791}
792
793/// Walk a rule at a vertex inside a μ-binder. The wrapping frame is
794/// pushed before recursion and popped after, so any SYMBOL inside
795/// `rule` that re-enters the same `(vertex_id, rule_name)` pair
796/// returns the empty sequence (μ X . body unfolds once).
797fn walk_in_mu_frame(
798    protocol: &str,
799    schema: &Schema,
800    grammar: &Grammar,
801    vertex_id: &panproto_gat::Name,
802    rule_name: &str,
803    rule: &Production,
804    cursor: &mut ChildCursor<'_>,
805    out: &mut Output<'_>,
806) -> Result<(), ParseError> {
807    let key = (vertex_id.to_string(), rule_name.to_owned());
808    let inserted = EMIT_MU_FRAMES.with(|frames| frames.borrow_mut().insert(key.clone()));
809    if !inserted {
810        // We are already walking this rule at this vertex deeper in
811        // the call stack. The coinductive μ-fixed-point reading
812        // returns the empty sequence here; the surrounding
813        // production resumes after the SYMBOL.
814        return Ok(());
815    }
816    let result = emit_production(protocol, schema, grammar, vertex_id, rule, cursor, out);
817    EMIT_MU_FRAMES.with(|frames| {
818        frames.borrow_mut().remove(&key);
819    });
820    result
821}
822
823fn emit_production(
824    protocol: &str,
825    schema: &Schema,
826    grammar: &Grammar,
827    vertex_id: &panproto_gat::Name,
828    production: &Production,
829    cursor: &mut ChildCursor<'_>,
830    out: &mut Output<'_>,
831) -> Result<(), ParseError> {
832    let depth = EMIT_DEPTH.with(|d| {
833        let v = d.get() + 1;
834        d.set(v);
835        v
836    });
837    if depth > 500 {
838        EMIT_DEPTH.with(|d| d.set(d.get() - 1));
839        return Err(ParseError::EmitFailed {
840            protocol: protocol.to_owned(),
841            reason: format!(
842                "emit_production recursion >500 (likely a cyclic grammar; \
843                     vertex='{vertex_id}')"
844            ),
845        });
846    }
847    let result = emit_production_inner(
848        protocol, schema, grammar, vertex_id, production, cursor, out,
849    );
850    EMIT_DEPTH.with(|d| d.set(d.get() - 1));
851    result
852}
853
854fn emit_production_inner(
855    protocol: &str,
856    schema: &Schema,
857    grammar: &Grammar,
858    vertex_id: &panproto_gat::Name,
859    production: &Production,
860    cursor: &mut ChildCursor<'_>,
861    out: &mut Output<'_>,
862) -> Result<(), ParseError> {
863    match production {
864        Production::String { value } => {
865            out.token(value);
866            Ok(())
867        }
868        Production::Pattern { value } => {
869            if let Some(literal) = literal_value(schema, vertex_id) {
870                out.token(literal);
871            } else {
872                out.token(&placeholder_for_pattern(value));
873            }
874            Ok(())
875        }
876        Production::Blank => Ok(()),
877        Production::Symbol { name } => {
878            // Inside a FIELD body, a SYMBOL consumes by field-name on
879            // the outer cursor rather than searching by symbol-match.
880            // This covers the simple `FIELD(SYMBOL X)` case as well as
881            // every nesting under FIELD that contains SYMBOLs (SEQ,
882            // CHOICE, REPEAT, OPTIONAL, ALIAS). Without the override,
883            // shapes like `field('args', commaSep1($.X))` consume one
884            // field edge in the FIELD handler and then the REPEAT
885            // inside SEQ searches the consumed child's cursor — where
886            // no sibling field edges sit — and breaks after one
887            // iteration.
888            if let Some(field) = current_field_context() {
889                if let Some(edge) = cursor.take_field(&field) {
890                    return emit_in_child_context(
891                        protocol, schema, grammar, &edge.tgt, production, out,
892                    );
893                }
894                // No matching field-named edge left on the outer
895                // cursor. Surface nothing; the surrounding REPEAT /
896                // OPTIONAL / CHOICE backtracks the literal tokens it
897                // emitted on this iteration when it sees no progress.
898                return Ok(());
899            }
900            if name.starts_with('_') {
901                // Hidden rule: not a vertex kind on the schema side.
902                // Inline-expand the rule body so its children take
903                // edges from the current cursor, instead of trying to
904                // take a single child edge that "satisfies" the
905                // hidden rule and discarding the rest of the body
906                // (which would drop tokens like `=` and the trailing
907                // value SYMBOL inside e.g. TOML's `_inline_pair`).
908                //
909                // Wrapped in a μ-frame so a hidden rule that
910                // references its own kind cyclically (or another
911                // hidden rule that closes the cycle) unfolds once
912                // and then collapses to the empty sequence at the
913                // second visit, rather than blowing the stack.
914                if let Some(rule) = grammar.rules.get(name) {
915                    walk_in_mu_frame(
916                        protocol, schema, grammar, vertex_id, name, rule, cursor, out,
917                    )
918                } else {
919                    // External hidden rule (declared in the
920                    // grammar's `externals` block, scanned by C code,
921                    // not listed in `rules`). Heuristic fallback by
922                    // name:
923                    //
924                    // - `_indent` / `*_indent`: open an indent block.
925                    //   Indent-based grammars (Python, YAML, qvr)
926                    //   declare an `_indent` external scanner before
927                    //   the body of a block-bodied declaration; the
928                    //   emitted output is unparseable without the
929                    //   corresponding indentation jump.
930                    // - `_dedent` / `*_dedent`: close the matching
931                    //   indent block.
932                    // - `_newline` / `*_line_ending` / `*_or_eof`:
933                    //   universally newline-or-empty; emitting a
934                    //   single newline is the right default for
935                    //   grammars like TOML whose `pair` SEQ trails
936                    //   into `_line_ending_or_eof`.
937                    //
938                    // Anything else falls through silently — better
939                    // to drop an unknown external token than to
940                    // invent one that confuses re-parsing.
941                    if name == "_indent" || name.ends_with("_indent") {
942                        out.indent_open();
943                    } else if name == "_dedent" || name.ends_with("_dedent") {
944                        out.indent_close();
945                    } else if name.contains("line_ending")
946                        || name.contains("newline")
947                        || name.ends_with("_or_eof")
948                    {
949                        out.newline();
950                    }
951                    Ok(())
952                }
953            } else if let Some(edge) = take_symbol_match(grammar, schema, cursor, name) {
954                // For supertype / hidden-rule dispatch the child's
955                // own kind names the actual production to walk
956                // (`child.kind` IS the subtype). For ALIAS the
957                // dependent-optic context is carried by the
958                // surrounding `Production::Alias` branch, which calls
959                // `emit_aliased_child` directly; we don't reach here
960                // for that case. So walking `grammar.rules[child.kind]`
961                // via `emit_vertex` is correct: the dependent-optic
962                // path is preserved at every site where it actually
963                // diverges from `child.kind`.
964                emit_vertex(protocol, schema, grammar, &edge.tgt, out)
965            } else if vertex_id_kind(schema, vertex_id) == Some(name.as_str()) {
966                let rule = grammar
967                    .rules
968                    .get(name)
969                    .ok_or_else(|| ParseError::EmitFailed {
970                        protocol: protocol.to_owned(),
971                        reason: format!("no production for SYMBOL '{name}'"),
972                    })?;
973                // Self-reference (`X = ... SYMBOL X ...`): wrap in a
974                // μ-frame so re-entry collapses to the empty sequence.
975                walk_in_mu_frame(
976                    protocol, schema, grammar, vertex_id, name, rule, cursor, out,
977                )
978            } else {
979                // Named rule with no matching child: emit nothing and
980                // let the surrounding CHOICE / OPTIONAL / REPEAT
981                // resolve the absence.
982                Ok(())
983            }
984        }
985        Production::Seq { members } => {
986            for member in members {
987                emit_production(protocol, schema, grammar, vertex_id, member, cursor, out)?;
988            }
989            Ok(())
990        }
991        Production::Choice { members } => {
992            if let Some(matched) =
993                pick_choice_with_cursor(schema, grammar, vertex_id, cursor, members)
994            {
995                emit_production(protocol, schema, grammar, vertex_id, matched, cursor, out)
996            } else {
997                Ok(())
998            }
999        }
1000        Production::Repeat { content } | Production::Repeat1 { content } => {
1001            let mut emitted_any = false;
1002            loop {
1003                let cursor_snap = cursor.consumed.clone();
1004                let out_snap = out.snapshot();
1005                let consumed_before = cursor.consumed.iter().filter(|&&c| c).count();
1006                let result =
1007                    emit_production(protocol, schema, grammar, vertex_id, content, cursor, out);
1008                let consumed_after = cursor.consumed.iter().filter(|&&c| c).count();
1009                if result.is_err() || consumed_after == consumed_before {
1010                    cursor.consumed = cursor_snap;
1011                    out.restore(out_snap);
1012                    break;
1013                }
1014                emitted_any = true;
1015            }
1016            if matches!(production, Production::Repeat1 { .. }) && !emitted_any {
1017                emit_production(protocol, schema, grammar, vertex_id, content, cursor, out)?;
1018            }
1019            Ok(())
1020        }
1021        Production::Optional { content } => {
1022            let cursor_snap = cursor.consumed.clone();
1023            let out_snap = out.snapshot();
1024            let consumed_before = cursor.consumed.iter().filter(|&&c| c).count();
1025            let result =
1026                emit_production(protocol, schema, grammar, vertex_id, content, cursor, out);
1027            // OPTIONAL is a backtracking site: if the inner production
1028            // errored *or* made no progress without leaving a witness
1029            // constraint, restore both cursor and output to their
1030            // pre-attempt state. Mirrors `Repeat`'s loop body.
1031            if result.is_err() {
1032                cursor.consumed = cursor_snap;
1033                out.restore(out_snap);
1034                return result;
1035            }
1036            let consumed_after = cursor.consumed.iter().filter(|&&c| c).count();
1037            if consumed_after == consumed_before
1038                && !has_relevant_constraint(content, schema, vertex_id)
1039            {
1040                cursor.consumed = cursor_snap;
1041                out.restore(out_snap);
1042            }
1043            Ok(())
1044        }
1045        Production::Field { name, content } => {
1046            // Set the field context for the duration of `content`'s
1047            // walk and emit the content against the *outer* cursor.
1048            // The SYMBOL handler picks up the context and pulls
1049            // successive `take_field(name)` edges as it encounters
1050            // SYMBOLs anywhere under `content` (under SEQ, CHOICE,
1051            // REPEAT, OPTIONAL, ALIAS — arbitrarily nested). This
1052            // subsumes the prior carve-outs for FIELD(REPEAT(...)),
1053            // FIELD(REPEAT1(...)), and the bare FIELD(SYMBOL ...)
1054            // case, and adds coverage for
1055            // `field('xs', commaSep1($.X))` which expands to
1056            // FIELD(SEQ(SYMBOL X, REPEAT(SEQ(',', SYMBOL X)))) and
1057            // any other shape where REPEAT/REPEAT1 sits inside SEQ /
1058            // CHOICE / OPTIONAL under a FIELD. A FIELD that wraps a
1059            // non-SYMBOL production (e.g. `field('op', '+')` or
1060            // `field('op', CHOICE(STRING ...))`) still works: STRING
1061            // handlers ignore the context and emit literals
1062            // directly, so the operator token survives the round
1063            // trip.
1064            let _guard = push_field_context(name);
1065            emit_production(protocol, schema, grammar, vertex_id, content, cursor, out)
1066        }
1067        Production::Alias {
1068            content,
1069            named,
1070            value,
1071        } => {
1072            // A named ALIAS rewrites the parser-visible kind to
1073            // `value`. If the cursor has an unconsumed child whose
1074            // kind matches that alias name, take it and emit the
1075            // child using the alias's INNER content as the rule
1076            // (e.g. `ALIAS { SYMBOL real_rule, value: "kind_x" }`
1077            // means a `kind_x` vertex on the schema should be walked
1078            // through `real_rule`'s body, not through whatever rule
1079            // happens to be keyed under `kind_x`). This is the
1080            // dependent-optic shape: the rule the emitter walks at a
1081            // child position is determined by the parent's chosen
1082            // alias, not by the child kind alone — without it,
1083            // grammars like YAML that introduce the same kind through
1084            // many ALIAS sites lose the parent context the moment
1085            // emit_vertex is called.
1086            if *named && !value.is_empty() {
1087                if let Some(edge) = cursor.take_matching(|edge| {
1088                    schema
1089                        .vertices
1090                        .get(&edge.tgt)
1091                        .map(|v| v.kind.as_ref() == value.as_str())
1092                        .unwrap_or(false)
1093                }) {
1094                    return emit_aliased_child(protocol, schema, grammar, &edge.tgt, content, out);
1095                }
1096            }
1097            emit_production(protocol, schema, grammar, vertex_id, content, cursor, out)
1098        }
1099        Production::Token { content }
1100        | Production::ImmediateToken { content }
1101        | Production::Prec { content, .. }
1102        | Production::PrecLeft { content, .. }
1103        | Production::PrecRight { content, .. }
1104        | Production::PrecDynamic { content, .. }
1105        | Production::Reserved { content, .. } => {
1106            emit_production(protocol, schema, grammar, vertex_id, content, cursor, out)
1107        }
1108    }
1109}
1110
1111/// Take the next cursor edge whose target vertex's kind matches the
1112/// SYMBOL `name` directly or via inline expansion of a hidden rule.
1113fn take_symbol_match<'a>(
1114    grammar: &Grammar,
1115    schema: &Schema,
1116    cursor: &mut ChildCursor<'a>,
1117    name: &str,
1118) -> Option<&'a Edge> {
1119    cursor.take_matching(|edge| {
1120        let target_kind = schema.vertices.get(&edge.tgt).map(|v| v.kind.as_ref());
1121        kind_satisfies_symbol(grammar, target_kind, name)
1122    })
1123}
1124
1125/// Decide whether a schema vertex of kind `target_kind` satisfies a
1126/// SYMBOL `name` reference in the grammar.
1127///
1128/// Operates as an O(1) lookup against the precomputed subtype
1129/// closure built at [`Grammar::from_bytes`]. The semantic content is
1130/// "K satisfies SYMBOL S iff K is reachable from S by walking the
1131/// grammar's hidden, supertype, and named-alias dispatch": this is
1132/// exactly the relation tree-sitter induces on `(parser-visible kind,
1133/// rule-position)` pairs.
1134fn kind_satisfies_symbol(grammar: &Grammar, target_kind: Option<&str>, name: &str) -> bool {
1135    let Some(target) = target_kind else {
1136        return false;
1137    };
1138    if target == name {
1139        return true;
1140    }
1141    grammar
1142        .subtypes
1143        .get(target)
1144        .is_some_and(|set| set.contains(name))
1145}
1146
1147/// Emit a child reached through an ALIAS production using the
1148/// alias's inner content as the rule, not `grammar.rules[child.kind]`.
1149///
1150/// This carries the dependent-optic context across the ALIAS edge:
1151/// at the parent rule's site we know which underlying production the
1152/// alias wraps (typically `SYMBOL real_rule`), and that's the
1153/// production that should drive the emit walk on the child's
1154/// children. Looking up `grammar.rules.get(child.kind)` instead would
1155/// either fail (the renamed kind has no top-level rule, e.g. YAML's
1156/// `block_mapping_pair`) or pick an arbitrary same-kinded rule from
1157/// elsewhere in the grammar.
1158///
1159/// Walk-context invariant. The dependent-optic shape of `emit_pretty`
1160/// says: the production walked at any vertex is determined by the
1161/// path from the root through the grammar, not by the vertex kind in
1162/// isolation. Two dispatch sites realise that invariant:
1163///
1164/// * [`emit_vertex`] looks up `grammar.rules[child.kind]` and walks
1165///   it. Correct for supertype / hidden-rule dispatch: the child's
1166///   kind on the schema IS the subtype tree-sitter selected, so its
1167///   top-level rule is the right production to walk.
1168/// * `emit_aliased_child` threads the parent rule's `Production`
1169///   directly (the inner `content` of `Production::Alias`) and walks
1170///   it on the child's children. Correct for ALIAS dispatch: the
1171///   child's kind on the schema is the alias's `value` (a renamed
1172///   kind that may have no top-level rule), and the production to
1173///   walk is the alias's content body, supplied by the parent.
1174///
1175/// Together these cover every site where the rule-walked-at-child
1176/// diverges from `grammar.rules[child.kind]`; the recursion site for
1177/// plain SYMBOL therefore correctly delegates to `emit_vertex`, and
1178/// we do not need a richer `WalkContext` value passed by reference.
1179/// The grammar dependency is the thread.
1180fn emit_aliased_child(
1181    protocol: &str,
1182    schema: &Schema,
1183    grammar: &Grammar,
1184    child_id: &panproto_gat::Name,
1185    content: &Production,
1186    out: &mut Output<'_>,
1187) -> Result<(), ParseError> {
1188    // Leaf shortcut: if the child has a literal-value and no
1189    // structural children, emit the captured text. Identifiers and
1190    // similar terminals reach here when an ALIAS wraps a SYMBOL that
1191    // resolves to a PATTERN.
1192    if let Some(literal) = literal_value(schema, child_id) {
1193        if children_for(schema, child_id).is_empty() {
1194            out.token(literal);
1195            return Ok(());
1196        }
1197    }
1198
1199    // Resolve `content` to a rule when it's a SYMBOL (the dominant
1200    // shape: `ALIAS { content: SYMBOL real_rule, value: "kind_x" }`).
1201    if let Production::Symbol { name } = content {
1202        if let Some(rule) = grammar.rules.get(name) {
1203            let edges = children_for(schema, child_id);
1204            let mut cursor = ChildCursor::new(&edges);
1205            return emit_production(protocol, schema, grammar, child_id, rule, &mut cursor, out);
1206        }
1207    }
1208
1209    // Other ALIAS contents (CHOICE, SEQ, literals) walk in place.
1210    let edges = children_for(schema, child_id);
1211    let mut cursor = ChildCursor::new(&edges);
1212    emit_production(
1213        protocol,
1214        schema,
1215        grammar,
1216        child_id,
1217        content,
1218        &mut cursor,
1219        out,
1220    )
1221}
1222
1223fn emit_in_child_context(
1224    protocol: &str,
1225    schema: &Schema,
1226    grammar: &Grammar,
1227    child_id: &panproto_gat::Name,
1228    production: &Production,
1229    out: &mut Output<'_>,
1230) -> Result<(), ParseError> {
1231    // The child walks under its own production tree, with its own
1232    // FIELDs setting their own contexts. Clear the outer FIELD hint
1233    // so it does not leak through and cause sibling SYMBOLs inside
1234    // the child's body to mistakenly pull edges from the child's
1235    // cursor by the parent's field name.
1236    let _guard = clear_field_context();
1237    // If `production` is a structural wrapper (CHOICE / SEQ /
1238    // OPTIONAL / ...) whose referenced symbols cover the child's own
1239    // kind, the child IS the production's target node and the right
1240    // emit path is `emit_vertex(child)` (which honours the
1241    // literal-value leaf shortcut). Without this guard, FIELD(pattern,
1242    // CHOICE { _pattern, self }) on an identifier child walks the
1243    // CHOICE on the identifier's empty cursor, falls through to the
1244    // first non-BLANK alt, and loses the captured identifier text.
1245    if !matches!(production, Production::Symbol { .. }) {
1246        let child_kind = schema.vertices.get(child_id).map(|v| v.kind.as_ref());
1247        let symbols = referenced_symbols(production);
1248        if symbols
1249            .iter()
1250            .any(|s| kind_satisfies_symbol(grammar, child_kind, s) || child_kind == Some(s))
1251        {
1252            return emit_vertex(protocol, schema, grammar, child_id, out);
1253        }
1254    }
1255    match production {
1256        Production::Symbol { .. } => emit_vertex(protocol, schema, grammar, child_id, out),
1257        _ => {
1258            let edges = children_for(schema, child_id);
1259            let mut cursor = ChildCursor::new(&edges);
1260            emit_production(
1261                protocol,
1262                schema,
1263                grammar,
1264                child_id,
1265                production,
1266                &mut cursor,
1267                out,
1268            )
1269        }
1270    }
1271}
1272
1273fn pick_choice_with_cursor<'a>(
1274    schema: &Schema,
1275    grammar: &Grammar,
1276    vertex_id: &panproto_gat::Name,
1277    cursor: &ChildCursor<'_>,
1278    alternatives: &'a [Production],
1279) -> Option<&'a Production> {
1280    // Discriminator-driven dispatch (highest priority): when the
1281    // walker recorded a `chose-alt-fingerprint` constraint at parse
1282    // time, dispatch directly against that. This is the categorical
1283    // discriminator: it survives stripping of byte-position
1284    // constraints (so by-construction round-trips work) and is the
1285    // explicit witness of which CHOICE alternative the parser took.
1286    //
1287    // Falls back to the live `interstitial-*` substring blob when no
1288    // fingerprint is present (e.g. instances built by callers that
1289    // bypass the AstWalker). Both blobs are scored by the longest
1290    // STRING-literal token in an alternative that matches; the
1291    // length tiebreak prefers `&&` over `&`, `==` over `=`, etc.
1292    let constraint_blob = schema
1293        .constraints
1294        .get(vertex_id)
1295        .map(|cs| {
1296            let fingerprint: Option<&str> = cs
1297                .iter()
1298                .find(|c| c.sort.as_ref() == "chose-alt-fingerprint")
1299                .map(|c| c.value.as_str());
1300            if let Some(fp) = fingerprint {
1301                fp.to_owned()
1302            } else {
1303                cs.iter()
1304                    .filter(|c| {
1305                        let s = c.sort.as_ref();
1306                        s.starts_with("interstitial-") && !s.ends_with("-start-byte")
1307                    })
1308                    .map(|c| c.value.as_str())
1309                    .collect::<Vec<&str>>()
1310                    .join(" ")
1311            }
1312        })
1313        .unwrap_or_default();
1314    let child_kinds: Vec<&str> = schema
1315        .constraints
1316        .get(vertex_id)
1317        .and_then(|cs| {
1318            cs.iter()
1319                .find(|c| c.sort.as_ref() == "chose-alt-child-kinds")
1320                .map(|c| c.value.split_whitespace().collect())
1321        })
1322        .unwrap_or_default();
1323    if !constraint_blob.is_empty() {
1324        // Primary score: literal-token match length. This dominates
1325        // alt selection so existing language tests that depend on
1326        // literal-only fingerprints keep working.
1327        // Secondary score (tiebreaker only): named-symbol kind match
1328        // count, read from the separate `chose-alt-child-kinds`
1329        // constraint (kept apart from the literal fingerprint so
1330        // identifiers like `:` in the kind list don't contaminate the
1331        // literal match). An alt that matches the recorded kinds is a
1332        // stronger witness than one whose only
1333        // overlap is literal punctuation.
1334        let mut best_literal: usize = 0;
1335        let mut best_symbols: usize = 0;
1336        let mut best_alt: Option<&Production> = None;
1337        let mut tied = false;
1338        for alt in alternatives {
1339            let strings = literal_strings(alt);
1340            if strings.is_empty() {
1341                continue;
1342            }
1343            let literal_score = strings
1344                .iter()
1345                .filter(|s| constraint_blob.contains(s.as_str()))
1346                .map(String::len)
1347                .sum::<usize>();
1348            if literal_score == 0 {
1349                continue;
1350            }
1351            // Symbol score is computed only as a tiebreaker among alts
1352            // whose literal-token coverage is the same; it never lifts
1353            // an alt above one with a strictly higher literal score.
1354            // Reads the `chose-alt-child-kinds` constraint (a separate
1355            // sequence the walker emits, kept apart from the literal
1356            // fingerprint to avoid cross-contamination).
1357            let symbol_score = if literal_score >= best_literal && !child_kinds.is_empty() {
1358                let symbols = referenced_symbols(alt);
1359                symbols
1360                    .iter()
1361                    .filter(|sym| {
1362                        let sym_str: &str = sym;
1363                        if child_kinds.contains(&sym_str) {
1364                            return true;
1365                        }
1366                        grammar.subtypes.get(sym_str).is_some_and(|sub_set| {
1367                            sub_set
1368                                .iter()
1369                                .any(|sub| child_kinds.contains(&sub.as_str()))
1370                        })
1371                    })
1372                    .count()
1373            } else {
1374                0
1375            };
1376            let better = literal_score > best_literal
1377                || (literal_score == best_literal && symbol_score > best_symbols);
1378            let same = literal_score == best_literal && symbol_score == best_symbols;
1379            if better {
1380                best_literal = literal_score;
1381                best_symbols = symbol_score;
1382                best_alt = Some(alt);
1383                tied = false;
1384            } else if same && best_alt.is_some() {
1385                tied = true;
1386            }
1387        }
1388        // Only commit to an alt when the fingerprint discriminates it
1389        // uniquely. A tie means the alts share the same literal token
1390        // set (e.g. JSON's `string = CHOICE { SEQ { '"', '"' }, SEQ {
1391        // '"', _string_content, '"' } }` — both alts contain just the
1392        // two `"` tokens). In that case fall through to cursor-based
1393        // dispatch, which uses the actual edge structure.
1394        if let Some(alt) = best_alt {
1395            if !tied {
1396                return Some(alt);
1397            }
1398        }
1399    }
1400
1401    // Cursor-driven dispatch: pick the alternative whose body
1402    // references a SYMBOL covering the *first unconsumed* edge in
1403    // cursor order. `referenced_symbols` walks the alternative
1404    // recursively (across nested SEQs, REPEATs, OPTIONALs, FIELDs,
1405    // etc.) so a leading optional like `attribute_item` does not
1406    // block matching when only the trailing required symbol is
1407    // present on the schema.
1408    //
1409    // Ordering by the first unconsumed edge (rather than picking any
1410    // alternative whose SYMBOL set intersects the unconsumed
1411    // multiset) is what preserves schema edge order under
1412    // REPEAT(CHOICE(...)) productions. Without this rule, alt order
1413    // in the grammar's CHOICE determines the emission order, and a
1414    // schema with interleaved kinds like `[symbol, punct, int,
1415    // symbol, punct, int]` re-fuses to `[symbol, symbol, punct,
1416    // punct, int, int]` when emitted then re-parsed. The fix is the
1417    // categorical reading of REPEAT-over-list (list-shaped fold)
1418    // rather than REPEAT-over-multiset (unordered fold).
1419    let first_unconsumed_kind: Option<&str> = cursor
1420        .edges
1421        .iter()
1422        .enumerate()
1423        .find(|(i, _)| !cursor.consumed[*i])
1424        .and_then(|(_, edge)| schema.vertices.get(&edge.tgt).map(|v| v.kind.as_ref()));
1425    if let Some(target_kind) = first_unconsumed_kind {
1426        for alt in alternatives {
1427            let symbols = referenced_symbols(alt);
1428            if !symbols.is_empty()
1429                && symbols
1430                    .iter()
1431                    .any(|s| kind_satisfies_symbol(grammar, Some(target_kind), s))
1432            {
1433                return Some(alt);
1434            }
1435        }
1436    }
1437
1438    // FIELD dispatch: pick an alternative whose FIELD name matches an
1439    // unconsumed edge kind.
1440    let edge_kinds: Vec<&str> = cursor
1441        .edges
1442        .iter()
1443        .enumerate()
1444        .filter(|(i, _)| !cursor.consumed[*i])
1445        .map(|(_, e)| e.kind.as_ref())
1446        .collect();
1447    for alt in alternatives {
1448        if has_field_in(alt, &edge_kinds) {
1449            return Some(alt);
1450        }
1451    }
1452
1453    // No cursor-driven match. Fall back to:
1454    //
1455    // - BLANK (the explicit empty alternative) when present, so an
1456    //   OPTIONAL-shaped CHOICE compiles to nothing.
1457    // - The first non-`BLANK` alternative as a last resort, used by
1458    //   STRING-only alternatives (keyword tokens) and other choices
1459    //   that don't reach the cursor.
1460    //
1461    // The previous "match own_kind" branch is intentionally absent:
1462    // when an alt's first SYMBOL equals the current vertex's kind, the
1463    // caller is already emitting that vertex's own rule. Recursing
1464    // into the alt would cause a self-loop in the rule walk.
1465    let _ = (schema, vertex_id);
1466    if alternatives.iter().any(|a| matches!(a, Production::Blank)) {
1467        return alternatives.iter().find(|a| matches!(a, Production::Blank));
1468    }
1469    alternatives
1470        .iter()
1471        .find(|alt| !matches!(alt, Production::Blank))
1472}
1473
1474/// Collect every literal STRING token directly inside `production`
1475/// (without descending into SYMBOLs / hidden rules). Used to score
1476/// CHOICE alternatives against the parent vertex's interstitials so
1477/// the right operator / keyword form is picked when the schema
1478/// preserves interstitial fragments from a prior parse.
1479fn literal_strings(production: &Production) -> Vec<String> {
1480    let mut out = Vec::new();
1481    fn walk(p: &Production, out: &mut Vec<String>) {
1482        match p {
1483            Production::String { value } if !value.is_empty() => {
1484                out.push(value.clone());
1485            }
1486            Production::Choice { members } | Production::Seq { members } => {
1487                for m in members {
1488                    walk(m, out);
1489                }
1490            }
1491            Production::Repeat { content }
1492            | Production::Repeat1 { content }
1493            | Production::Optional { content }
1494            | Production::Field { content, .. }
1495            | Production::Alias { content, .. }
1496            | Production::Token { content }
1497            | Production::ImmediateToken { content }
1498            | Production::Prec { content, .. }
1499            | Production::PrecLeft { content, .. }
1500            | Production::PrecRight { content, .. }
1501            | Production::PrecDynamic { content, .. }
1502            | Production::Reserved { content, .. } => walk(content, out),
1503            _ => {}
1504        }
1505    }
1506    walk(production, &mut out);
1507    out
1508}
1509
1510/// Collect every SYMBOL name reachable from `production` without
1511/// crossing into nested rules. Used by `pick_choice_with_cursor` to
1512/// rank alternatives by "any SYMBOL inside this alt matches something
1513/// on the cursor", instead of just the first SYMBOL: a leading
1514/// optional like `attribute_item` then `parameter` is otherwise
1515/// rejected when only the parameter children are present.
1516fn referenced_symbols(production: &Production) -> Vec<&str> {
1517    let mut out = Vec::new();
1518    fn walk<'a>(p: &'a Production, out: &mut Vec<&'a str>) {
1519        match p {
1520            Production::Symbol { name } => out.push(name.as_str()),
1521            Production::Choice { members } | Production::Seq { members } => {
1522                for m in members {
1523                    walk(m, out);
1524                }
1525            }
1526            Production::Alias {
1527                content,
1528                named,
1529                value,
1530            } => {
1531                // A named ALIAS produces a child vertex whose kind is
1532                // the alias `value` (e.g. `ALIAS { content: STRING "=",
1533                // value: "punctuation", named: true }` introduces a
1534                // `punctuation` child). For cursor-driven dispatch to
1535                // recognise alts that emit such children, yield the
1536                // alias value as a referenced symbol. Anonymous aliases
1537                // do not introduce a named node and only need their
1538                // inner content's symbols.
1539                if *named && !value.is_empty() {
1540                    out.push(value.as_str());
1541                }
1542                walk(content, out);
1543            }
1544            Production::Repeat { content }
1545            | Production::Repeat1 { content }
1546            | Production::Optional { content }
1547            | Production::Field { content, .. }
1548            | Production::Token { content }
1549            | Production::ImmediateToken { content }
1550            | Production::Prec { content, .. }
1551            | Production::PrecLeft { content, .. }
1552            | Production::PrecRight { content, .. }
1553            | Production::PrecDynamic { content, .. }
1554            | Production::Reserved { content, .. } => walk(content, out),
1555            _ => {}
1556        }
1557    }
1558    walk(production, &mut out);
1559    out
1560}
1561
1562#[cfg(test)]
1563fn first_symbol(production: &Production) -> Option<&str> {
1564    match production {
1565        Production::Symbol { name } => Some(name),
1566        Production::Seq { members } => members.iter().find_map(first_symbol),
1567        Production::Choice { members } => members.iter().find_map(first_symbol),
1568        Production::Repeat { content }
1569        | Production::Repeat1 { content }
1570        | Production::Optional { content }
1571        | Production::Field { content, .. }
1572        | Production::Alias { content, .. }
1573        | Production::Token { content }
1574        | Production::ImmediateToken { content }
1575        | Production::Prec { content, .. }
1576        | Production::PrecLeft { content, .. }
1577        | Production::PrecRight { content, .. }
1578        | Production::PrecDynamic { content, .. }
1579        | Production::Reserved { content, .. } => first_symbol(content),
1580        _ => None,
1581    }
1582}
1583
1584fn has_field_in(production: &Production, edge_kinds: &[&str]) -> bool {
1585    match production {
1586        Production::Field { name, .. } => edge_kinds.contains(&name.as_str()),
1587        Production::Seq { members } | Production::Choice { members } => {
1588            members.iter().any(|m| has_field_in(m, edge_kinds))
1589        }
1590        Production::Repeat { content }
1591        | Production::Repeat1 { content }
1592        | Production::Optional { content }
1593        | Production::Alias { content, .. }
1594        | Production::Token { content }
1595        | Production::ImmediateToken { content }
1596        | Production::Prec { content, .. }
1597        | Production::PrecLeft { content, .. }
1598        | Production::PrecRight { content, .. }
1599        | Production::PrecDynamic { content, .. }
1600        | Production::Reserved { content, .. } => has_field_in(content, edge_kinds),
1601        _ => false,
1602    }
1603}
1604
1605fn has_relevant_constraint(
1606    production: &Production,
1607    schema: &Schema,
1608    vertex_id: &panproto_gat::Name,
1609) -> bool {
1610    let constraints = match schema.constraints.get(vertex_id) {
1611        Some(c) => c,
1612        None => return false,
1613    };
1614    fn walk(production: &Production, constraints: &[panproto_schema::Constraint]) -> bool {
1615        match production {
1616            Production::String { value } => constraints
1617                .iter()
1618                .any(|c| c.value == *value || c.sort.as_ref() == value),
1619            Production::Field { name, content } => {
1620                constraints.iter().any(|c| c.sort.as_ref() == name) || walk(content, constraints)
1621            }
1622            Production::Seq { members } | Production::Choice { members } => {
1623                members.iter().any(|m| walk(m, constraints))
1624            }
1625            Production::Repeat { content }
1626            | Production::Repeat1 { content }
1627            | Production::Optional { content }
1628            | Production::Alias { content, .. }
1629            | Production::Token { content }
1630            | Production::ImmediateToken { content }
1631            | Production::Prec { content, .. }
1632            | Production::PrecLeft { content, .. }
1633            | Production::PrecRight { content, .. }
1634            | Production::PrecDynamic { content, .. }
1635            | Production::Reserved { content, .. } => walk(content, constraints),
1636            _ => false,
1637        }
1638    }
1639    walk(production, constraints)
1640}
1641
1642fn children_for<'a>(schema: &'a Schema, vertex_id: &panproto_gat::Name) -> Vec<&'a Edge> {
1643    // Walk `outgoing` (insertion-ordered by SchemaBuilder via SmallVec
1644    // append) rather than the unordered `edges` HashMap so abstract
1645    // schemas under REPEAT(CHOICE(...)) preserve the order their edges
1646    // were inserted in. The previous implementation walked the HashMap
1647    // and sorted lexicographically by (kind, target id), which fused
1648    // interleaved children of the same kind into runs (e.g. a sequence
1649    // [symbol, punct, int, symbol, punct, int] became [symbol, symbol,
1650    // punct, punct, int, int] after the lex sort).
1651    let Some(edges) = schema.outgoing.get(vertex_id) else {
1652        return Vec::new();
1653    };
1654
1655    // Look up the canonical Edge reference (the key in `schema.edges`)
1656    // for each entry in `outgoing`. Falls back to the SmallVec entry if
1657    // the canonical key is missing, which would indicate index drift.
1658    let mut indexed: Vec<(usize, u32, &Edge)> = edges
1659        .iter()
1660        .enumerate()
1661        .map(|(i, e)| {
1662            let canonical = schema.edges.get_key_value(e).map_or(e, |(k, _)| k);
1663            let pos = schema.orderings.get(canonical).copied().unwrap_or(u32::MAX);
1664            (i, pos, canonical)
1665        })
1666        .collect();
1667
1668    // Stable sort by (explicit-ordering, insertion-index). Edges with
1669    // an explicit `orderings` entry come first in their declared order;
1670    // the remainder fall through in insertion order.
1671    indexed.sort_by_key(|(i, pos, _)| (*pos, *i));
1672    indexed.into_iter().map(|(_, _, e)| e).collect()
1673}
1674
1675fn vertex_id_kind<'a>(schema: &'a Schema, vertex_id: &panproto_gat::Name) -> Option<&'a str> {
1676    schema.vertices.get(vertex_id).map(|v| v.kind.as_ref())
1677}
1678
1679fn literal_value<'a>(schema: &'a Schema, vertex_id: &panproto_gat::Name) -> Option<&'a str> {
1680    schema
1681        .constraints
1682        .get(vertex_id)?
1683        .iter()
1684        .find(|c| c.sort.as_ref() == "literal-value")
1685        .map(|c| c.value.as_str())
1686}
1687
1688fn placeholder_for_pattern(pattern: &str) -> String {
1689    // Heuristic placeholder for unconstrained PATTERN terminals.
1690    //
1691    // First handle the "the regex IS a literal escape" cases that
1692    // tree-sitter grammars use as separators (`\n`, `\r\n`, `;`,
1693    // etc.); emitting the matching character is always preferable
1694    // to a `_x` identifier-like placeholder when the surrounding
1695    // grammar expects a separator.
1696    let simple_lit = decode_simple_pattern_literal(pattern);
1697    if let Some(lit) = simple_lit {
1698        return lit;
1699    }
1700
1701    if pattern.contains("[0-9]") || pattern.contains("\\d") {
1702        "0".into()
1703    } else if pattern.contains("[a-zA-Z_]") || pattern.contains("\\w") {
1704        "_x".into()
1705    } else if pattern.contains('"') || pattern.contains('\'') {
1706        "\"\"".into()
1707    } else {
1708        "_".into()
1709    }
1710}
1711
1712/// Decode a tree-sitter PATTERN whose regex is a simple literal
1713/// (newline, semicolon, comma, etc.) to the byte sequence it matches.
1714/// Returns `None` for patterns with character classes, alternations,
1715/// or quantifiers; the caller falls back to the heuristic placeholder.
1716fn decode_simple_pattern_literal(pattern: &str) -> Option<String> {
1717    // Skip patterns containing regex metachars that would broaden the
1718    // match beyond a single literal byte sequence.
1719    if pattern
1720        .chars()
1721        .any(|c| matches!(c, '[' | ']' | '(' | ')' | '*' | '+' | '?' | '|' | '{' | '}'))
1722    {
1723        return None;
1724    }
1725    let mut out = String::new();
1726    let mut chars = pattern.chars();
1727    while let Some(c) = chars.next() {
1728        if c == '\\' {
1729            match chars.next() {
1730                Some('n') => out.push('\n'),
1731                Some('r') => out.push('\r'),
1732                Some('t') => out.push('\t'),
1733                Some('\\') => out.push('\\'),
1734                Some('/') => out.push('/'),
1735                Some(other) => out.push(other),
1736                None => return None,
1737            }
1738        } else {
1739            out.push(c);
1740        }
1741    }
1742    Some(out)
1743}
1744
1745// ═══════════════════════════════════════════════════════════════════
1746// Token list output with Spacing algebra
1747// ═══════════════════════════════════════════════════════════════════
1748//
1749// Emit produces a free monoid over `Token`. Layout (spaces, newlines,
1750// indentation) is a homomorphism `Vec<Token> -> Vec<u8>` parameterised
1751// by `FormatPolicy`. Separating the structural output from the layout
1752// decision means each phase has one job: emit walks the grammar and
1753// pushes tokens; layout is a single fold, locally driven by adjacent
1754// pairs and a depth counter. Snapshot/restore is just `tokens.len()`.
1755
1756#[derive(Clone)]
1757enum Token {
1758    /// A user-visible terminal contributed by the grammar.
1759    Lit(String),
1760    /// `indent_open` marker emitted when a `Lit` matched the policy's
1761    /// open list. Carried as a separate token so layout can decide to
1762    /// break + indent without re-scanning.
1763    IndentOpen,
1764    /// `indent_close` marker emitted before a closer-`Lit`.
1765    IndentClose,
1766    /// "Break a line here if not already at line start" — used after
1767    /// statements/declarations and after open braces.
1768    LineBreak,
1769}
1770
1771struct Output<'a> {
1772    tokens: Vec<Token>,
1773    policy: &'a FormatPolicy,
1774}
1775
1776#[derive(Clone)]
1777struct OutputSnapshot {
1778    tokens_len: usize,
1779}
1780
1781impl<'a> Output<'a> {
1782    fn new(policy: &'a FormatPolicy) -> Self {
1783        Self {
1784            tokens: Vec::new(),
1785            policy,
1786        }
1787    }
1788
1789    fn token(&mut self, value: &str) {
1790        if value.is_empty() {
1791            return;
1792        }
1793
1794        if self.policy.indent_close.iter().any(|t| t == value) {
1795            self.tokens.push(Token::IndentClose);
1796        }
1797
1798        self.tokens.push(Token::Lit(value.to_owned()));
1799
1800        if self.policy.indent_open.iter().any(|t| t == value) {
1801            self.tokens.push(Token::IndentOpen);
1802            self.tokens.push(Token::LineBreak);
1803        } else if self.policy.line_break_after.iter().any(|t| t == value) {
1804            self.tokens.push(Token::LineBreak);
1805        }
1806    }
1807
1808    fn newline(&mut self) {
1809        self.tokens.push(Token::LineBreak);
1810    }
1811
1812    /// Open an indent scope: subsequent `LineBreak`s render at the
1813    /// new depth until a matching `indent_close` pops it. Used by the
1814    /// external-token fallback to render indent-based grammars'
1815    /// `_indent` scanner outputs.
1816    fn indent_open(&mut self) {
1817        self.tokens.push(Token::IndentOpen);
1818        self.tokens.push(Token::LineBreak);
1819    }
1820
1821    /// Close one indent scope opened by `indent_open`.
1822    fn indent_close(&mut self) {
1823        self.tokens.push(Token::IndentClose);
1824    }
1825
1826    fn snapshot(&self) -> OutputSnapshot {
1827        OutputSnapshot {
1828            tokens_len: self.tokens.len(),
1829        }
1830    }
1831
1832    fn restore(&mut self, snap: OutputSnapshot) {
1833        self.tokens.truncate(snap.tokens_len);
1834    }
1835
1836    fn finish(self) -> Vec<u8> {
1837        layout(&self.tokens, self.policy)
1838    }
1839}
1840
1841/// Fold a token list into bytes. The algebra:
1842/// * adjacent `Lit`s get a single space iff `needs_space_between(a, b)`,
1843/// * `IndentOpen` / `IndentClose` adjust a depth counter,
1844/// * `LineBreak` writes `\n` if not already at line start, then the
1845///   next `Lit` writes `indent * indent_width` spaces of indent.
1846fn layout(tokens: &[Token], policy: &FormatPolicy) -> Vec<u8> {
1847    let mut bytes = Vec::new();
1848    let mut indent: usize = 0;
1849    let mut at_line_start = true;
1850    let mut last_lit: Option<&str> = None;
1851    // True iff, at the moment `last_lit` was emitted, the cursor was at a
1852    // position where the grammar expects an operand: start of stream / line,
1853    // just after an open paren / bracket / brace, just after a separator like
1854    // `,` or `;`, or just after a binary / assignment operator. Used by
1855    // `needs_space_between` to recognise `last_lit` as a tight unary prefix
1856    // (`f(-1.0)`) rather than a spaced binary operator (`a - b`).
1857    let mut last_was_in_operand_position = true;
1858    let mut expecting_operand = true;
1859    let newline = policy.newline.as_bytes();
1860    let separator = policy.separator.as_bytes();
1861
1862    for tok in tokens {
1863        match tok {
1864            Token::IndentOpen => indent += 1,
1865            Token::IndentClose => {
1866                indent = indent.saturating_sub(1);
1867                if !at_line_start {
1868                    bytes.extend_from_slice(newline);
1869                    at_line_start = true;
1870                    expecting_operand = true;
1871                }
1872            }
1873            Token::LineBreak => {
1874                if !at_line_start {
1875                    bytes.extend_from_slice(newline);
1876                    at_line_start = true;
1877                    expecting_operand = true;
1878                }
1879            }
1880            Token::Lit(value) => {
1881                if at_line_start {
1882                    bytes.extend(std::iter::repeat_n(b' ', indent * policy.indent_width));
1883                } else if let Some(prev) = last_lit {
1884                    if needs_space_between(prev, value, last_was_in_operand_position) {
1885                        bytes.extend_from_slice(separator);
1886                    }
1887                }
1888                bytes.extend_from_slice(value.as_bytes());
1889                at_line_start = false;
1890                last_was_in_operand_position = expecting_operand;
1891                expecting_operand = leaves_operand_position(value);
1892                last_lit = Some(value.as_str());
1893            }
1894        }
1895    }
1896
1897    if !at_line_start {
1898        bytes.extend_from_slice(newline);
1899    }
1900    bytes
1901}
1902
1903/// True iff emitting `tok` leaves the cursor in a position where the
1904/// grammar expects an operand next. Operand-introducing tokens are open
1905/// punctuation, separators, and operator-like strings; operand-terminating
1906/// tokens are identifiers, literals, and closing punctuation.
1907fn leaves_operand_position(tok: &str) -> bool {
1908    if tok.is_empty() {
1909        return true;
1910    }
1911    if is_punct_open(tok) {
1912        return true;
1913    }
1914    if matches!(tok, "," | ";") {
1915        return true;
1916    }
1917    if is_punct_close(tok) {
1918        return false;
1919    }
1920    if first_is_alnum_or_underscore(tok) || last_ends_with_alnum(tok) {
1921        return false;
1922    }
1923    // Pure punctuation/operator runs (`=`, `+`, `-`, `<=`, `>>`, …) leave
1924    // the cursor expecting another operand.
1925    true
1926}
1927
1928fn needs_space_between(last: &str, next: &str, expecting_operand: bool) -> bool {
1929    if last.is_empty() || next.is_empty() {
1930        return false;
1931    }
1932    if is_punct_open(last) || is_punct_open(next) {
1933        return false;
1934    }
1935    if is_punct_close(next) {
1936        return false;
1937    }
1938    if is_punct_close(last) && is_punct_punctuation(next) {
1939        return false;
1940    }
1941    if last == "." || next == "." {
1942        return false;
1943    }
1944    // Tight unary prefix: `last` is a sign/logical-not operator emitted
1945    // where the grammar expected an operand, so it glues to `next`.
1946    // `expecting_operand` here means: just before `last` was emitted,
1947    // the cursor expected an operand, which makes `last` a unary prefix.
1948    // Examples: `f(-1.0)`, `[ -2, 3 ]`, `return -x`, `a = !flag`.
1949    if expecting_operand && is_unary_prefix_operator(last) && first_is_operand_start(next) {
1950        return false;
1951    }
1952    if last_is_word_like(last) && first_is_word_like(next) {
1953        return true;
1954    }
1955    if last_ends_with_alnum(last) && first_is_alnum_or_underscore(next) {
1956        return true;
1957    }
1958    // Adjacent operator runs: keep them apart so the lexer doesn't glue
1959    // `>` and `=` into `>=` unintentionally.
1960    true
1961}
1962
1963fn is_unary_prefix_operator(s: &str) -> bool {
1964    matches!(s, "-" | "+" | "!" | "~")
1965}
1966
1967fn first_is_operand_start(s: &str) -> bool {
1968    s.chars()
1969        .next()
1970        .map(|c| c.is_alphanumeric() || c == '_' || c == '.' || c == '(')
1971        .unwrap_or(false)
1972}
1973
1974fn is_punct_open(s: &str) -> bool {
1975    matches!(s, "(" | "[" | "{" | "\"" | "'" | "`")
1976}
1977
1978fn is_punct_close(s: &str) -> bool {
1979    matches!(s, ")" | "]" | "}" | "," | ";" | ":" | "\"" | "'" | "`")
1980}
1981
1982fn is_punct_punctuation(s: &str) -> bool {
1983    matches!(s, "," | ";" | ":" | "." | ")" | "]" | "}")
1984}
1985
1986fn last_is_word_like(s: &str) -> bool {
1987    s.chars()
1988        .next_back()
1989        .map(|c| c.is_alphanumeric() || c == '_')
1990        .unwrap_or(false)
1991}
1992
1993fn first_is_word_like(s: &str) -> bool {
1994    s.chars()
1995        .next()
1996        .map(|c| c.is_alphanumeric() || c == '_')
1997        .unwrap_or(false)
1998}
1999
2000fn last_ends_with_alnum(s: &str) -> bool {
2001    s.chars()
2002        .next_back()
2003        .map(char::is_alphanumeric)
2004        .unwrap_or(false)
2005}
2006
2007fn first_is_alnum_or_underscore(s: &str) -> bool {
2008    s.chars()
2009        .next()
2010        .map(|c| c.is_alphanumeric() || c == '_')
2011        .unwrap_or(false)
2012}
2013
2014#[cfg(test)]
2015mod tests {
2016    use super::*;
2017
2018    #[test]
2019    fn parses_simple_grammar_json() {
2020        let bytes = br#"{
2021            "name": "tiny",
2022            "rules": {
2023                "program": {
2024                    "type": "SEQ",
2025                    "members": [
2026                        {"type": "STRING", "value": "hello"},
2027                        {"type": "STRING", "value": ";"}
2028                    ]
2029                }
2030            }
2031        }"#;
2032        let g = Grammar::from_bytes("tiny", bytes).expect("valid tiny grammar");
2033        assert!(g.rules.contains_key("program"));
2034    }
2035
2036    #[test]
2037    fn output_emits_punctuation_without_leading_space() {
2038        let policy = FormatPolicy::default();
2039        let mut out = Output::new(&policy);
2040        out.token("foo");
2041        out.token("(");
2042        out.token(")");
2043        out.token(";");
2044        let bytes = out.finish();
2045        let s = std::str::from_utf8(&bytes).expect("ascii output");
2046        assert!(s.starts_with("foo();"), "got {s:?}");
2047    }
2048
2049    #[test]
2050    fn grammar_from_bytes_rejects_malformed_input() {
2051        let result = Grammar::from_bytes("malformed", b"not json");
2052        let err = result.expect_err("malformed bytes must yield Err");
2053        let msg = err.to_string();
2054        assert!(
2055            msg.contains("malformed"),
2056            "error message should name the protocol: {msg:?}"
2057        );
2058    }
2059
2060    #[test]
2061    fn output_indents_after_open_brace() {
2062        let policy = FormatPolicy::default();
2063        let mut out = Output::new(&policy);
2064        out.token("fn");
2065        out.token("foo");
2066        out.token("(");
2067        out.token(")");
2068        out.token("{");
2069        out.token("body");
2070        out.token("}");
2071        let bytes = out.finish();
2072        let s = std::str::from_utf8(&bytes).expect("ascii output");
2073        assert!(s.contains("{\n"), "newline after opening brace: {s:?}");
2074        assert!(s.contains("body"), "body inside block: {s:?}");
2075        assert!(s.ends_with("}\n"), "newline after closing brace: {s:?}");
2076    }
2077
2078    #[test]
2079    fn output_no_space_between_word_and_dot() {
2080        let policy = FormatPolicy::default();
2081        let mut out = Output::new(&policy);
2082        out.token("foo");
2083        out.token(".");
2084        out.token("bar");
2085        let bytes = out.finish();
2086        let s = std::str::from_utf8(&bytes).expect("ascii output");
2087        assert!(s.starts_with("foo.bar"), "no space around dot: {s:?}");
2088    }
2089
2090    #[test]
2091    fn output_snapshot_restore_truncates_bytes() {
2092        let policy = FormatPolicy::default();
2093        let mut out = Output::new(&policy);
2094        out.token("keep");
2095        let snap = out.snapshot();
2096        out.token("drop");
2097        out.token("more");
2098        out.restore(snap);
2099        out.token("after");
2100        let bytes = out.finish();
2101        let s = std::str::from_utf8(&bytes).expect("ascii output");
2102        assert!(s.contains("keep"), "kept token survives: {s:?}");
2103        assert!(s.contains("after"), "post-restore token visible: {s:?}");
2104        assert!(!s.contains("drop"), "rolled-back token removed: {s:?}");
2105        assert!(!s.contains("more"), "rolled-back token removed: {s:?}");
2106    }
2107
2108    #[test]
2109    fn child_cursor_take_field_consumes_once() {
2110        let edges_owned: Vec<Edge> = vec![Edge {
2111            src: panproto_gat::Name::from("p"),
2112            tgt: panproto_gat::Name::from("c"),
2113            kind: panproto_gat::Name::from("name"),
2114            name: None,
2115        }];
2116        let edges: Vec<&Edge> = edges_owned.iter().collect();
2117        let mut cursor = ChildCursor::new(&edges);
2118        let first = cursor.take_field("name");
2119        let second = cursor.take_field("name");
2120        assert!(first.is_some(), "first take returns the edge");
2121        assert!(
2122            second.is_none(),
2123            "second take returns None (already consumed)"
2124        );
2125    }
2126
2127    #[test]
2128    fn child_cursor_take_matching_predicate() {
2129        let edges_owned: Vec<Edge> = vec![
2130            Edge {
2131                src: "p".into(),
2132                tgt: "c1".into(),
2133                kind: "child_of".into(),
2134                name: None,
2135            },
2136            Edge {
2137                src: "p".into(),
2138                tgt: "c2".into(),
2139                kind: "key".into(),
2140                name: None,
2141            },
2142        ];
2143        let edges: Vec<&Edge> = edges_owned.iter().collect();
2144        let mut cursor = ChildCursor::new(&edges);
2145        assert!(cursor.has_matching(|e| e.kind.as_ref() == "key"));
2146        let taken = cursor.take_matching(|e| e.kind.as_ref() == "key");
2147        assert!(taken.is_some());
2148        assert!(
2149            !cursor.has_matching(|e| e.kind.as_ref() == "key"),
2150            "consumed edge no longer matches"
2151        );
2152        assert!(
2153            cursor.has_matching(|e| e.kind.as_ref() == "child_of"),
2154            "the other edge is still available"
2155        );
2156    }
2157
2158    #[test]
2159    fn kind_satisfies_symbol_direct_match() {
2160        let bytes = br#"{
2161            "name": "tiny",
2162            "rules": {
2163                "x": {"type": "STRING", "value": "x"}
2164            }
2165        }"#;
2166        let g = Grammar::from_bytes("tiny", bytes).expect("valid grammar");
2167        assert!(kind_satisfies_symbol(&g, Some("x"), "x"));
2168        assert!(!kind_satisfies_symbol(&g, Some("y"), "x"));
2169        assert!(!kind_satisfies_symbol(&g, None, "x"));
2170    }
2171
2172    #[test]
2173    fn kind_satisfies_symbol_through_hidden_rule() {
2174        let bytes = br#"{
2175            "name": "tiny",
2176            "rules": {
2177                "_value": {
2178                    "type": "CHOICE",
2179                    "members": [
2180                        {"type": "SYMBOL", "name": "object"},
2181                        {"type": "SYMBOL", "name": "number"}
2182                    ]
2183                },
2184                "object": {"type": "STRING", "value": "{}"},
2185                "number": {"type": "PATTERN", "value": "[0-9]+"}
2186            }
2187        }"#;
2188        let g = Grammar::from_bytes("tiny", bytes).expect("valid grammar");
2189        assert!(
2190            kind_satisfies_symbol(&g, Some("number"), "_value"),
2191            "number is reachable from _value via CHOICE"
2192        );
2193        assert!(
2194            kind_satisfies_symbol(&g, Some("object"), "_value"),
2195            "object is reachable from _value via CHOICE"
2196        );
2197        assert!(
2198            !kind_satisfies_symbol(&g, Some("string"), "_value"),
2199            "string is NOT among the alternatives"
2200        );
2201    }
2202
2203    #[test]
2204    fn first_symbol_skips_string_terminals() {
2205        let prod: Production = serde_json::from_str(
2206            r#"{
2207                "type": "SEQ",
2208                "members": [
2209                    {"type": "STRING", "value": "{"},
2210                    {"type": "SYMBOL", "name": "body"},
2211                    {"type": "STRING", "value": "}"}
2212                ]
2213            }"#,
2214        )
2215        .expect("valid SEQ");
2216        assert_eq!(first_symbol(&prod), Some("body"));
2217    }
2218
2219    #[test]
2220    fn placeholder_for_pattern_routes_by_regex_class() {
2221        assert_eq!(placeholder_for_pattern("[0-9]+"), "0");
2222        assert_eq!(placeholder_for_pattern("[a-zA-Z_]\\w*"), "_x");
2223        assert_eq!(placeholder_for_pattern("\"[^\"]*\""), "\"\"");
2224        assert_eq!(placeholder_for_pattern("\\d+\\.\\d+"), "0");
2225    }
2226
2227    #[test]
2228    fn format_policy_default_breaks_after_semicolon() {
2229        let policy = FormatPolicy::default();
2230        assert!(policy.line_break_after.iter().any(|t| t == ";"));
2231        assert!(policy.indent_open.iter().any(|t| t == "{"));
2232        assert!(policy.indent_close.iter().any(|t| t == "}"));
2233        assert_eq!(policy.indent_width, 2);
2234    }
2235
2236    #[test]
2237    fn placeholder_decodes_literal_pattern_separators() {
2238        // PATTERN regexes that match a single literal byte sequence
2239        // (newline, semicolon, comma) emit the bytes verbatim instead
2240        // of falling through to the `_` catch-all.
2241        assert_eq!(placeholder_for_pattern("\\n"), "\n");
2242        assert_eq!(placeholder_for_pattern("\\r\\n"), "\r\n");
2243        assert_eq!(placeholder_for_pattern(";"), ";");
2244        // Patterns with character classes / alternation still route
2245        // through the heuristic.
2246        assert_eq!(placeholder_for_pattern("[0-9]+"), "0");
2247        assert_eq!(placeholder_for_pattern("a|b"), "_");
2248    }
2249
2250    #[test]
2251    fn supertypes_decode_from_grammar_json_strings() {
2252        // Tree-sitter older grammars list supertypes as bare strings.
2253        let bytes = br#"{
2254            "name": "tiny",
2255            "supertypes": ["expression"],
2256            "rules": {
2257                "expression": {
2258                    "type": "CHOICE",
2259                    "members": [
2260                        {"type": "SYMBOL", "name": "binary_expression"},
2261                        {"type": "SYMBOL", "name": "identifier"}
2262                    ]
2263                },
2264                "binary_expression": {"type": "STRING", "value": "x"},
2265                "identifier": {"type": "PATTERN", "value": "[a-z]+"}
2266            }
2267        }"#;
2268        let g = Grammar::from_bytes("tiny", bytes).expect("parse");
2269        assert!(g.supertypes.contains("expression"));
2270        // identifier matches the supertype `expression`.
2271        assert!(kind_satisfies_symbol(&g, Some("identifier"), "expression"));
2272        // unrelated kinds do not.
2273        assert!(!kind_satisfies_symbol(&g, Some("string"), "expression"));
2274    }
2275
2276    #[test]
2277    fn supertypes_decode_from_grammar_json_objects() {
2278        // Recent grammars list supertypes as `{type: SYMBOL, name: ...}`
2279        // entries instead of bare strings.
2280        let bytes = br#"{
2281            "name": "tiny",
2282            "supertypes": [{"type": "SYMBOL", "name": "stmt"}],
2283            "rules": {
2284                "stmt": {
2285                    "type": "CHOICE",
2286                    "members": [
2287                        {"type": "SYMBOL", "name": "while_stmt"},
2288                        {"type": "SYMBOL", "name": "if_stmt"}
2289                    ]
2290                },
2291                "while_stmt": {"type": "STRING", "value": "while"},
2292                "if_stmt": {"type": "STRING", "value": "if"}
2293            }
2294        }"#;
2295        let g = Grammar::from_bytes("tiny", bytes).expect("parse");
2296        assert!(g.supertypes.contains("stmt"));
2297        assert!(kind_satisfies_symbol(&g, Some("while_stmt"), "stmt"));
2298    }
2299
2300    #[test]
2301    fn alias_value_matches_kind() {
2302        // A named ALIAS rewrites the parser-visible kind to `value`;
2303        // `kind_satisfies_symbol` should accept that rewritten kind
2304        // when looking up the original SYMBOL.
2305        let bytes = br#"{
2306            "name": "tiny",
2307            "rules": {
2308                "_package_identifier": {
2309                    "type": "ALIAS",
2310                    "named": true,
2311                    "value": "package_identifier",
2312                    "content": {"type": "SYMBOL", "name": "identifier"}
2313                },
2314                "identifier": {"type": "PATTERN", "value": "[a-z]+"}
2315            }
2316        }"#;
2317        let g = Grammar::from_bytes("tiny", bytes).expect("parse");
2318        assert!(kind_satisfies_symbol(
2319            &g,
2320            Some("package_identifier"),
2321            "_package_identifier"
2322        ));
2323    }
2324
2325    #[test]
2326    fn referenced_symbols_walks_nested_seq() {
2327        let prod: Production = serde_json::from_str(
2328            r#"{
2329                "type": "SEQ",
2330                "members": [
2331                    {"type": "CHOICE", "members": [
2332                        {"type": "SYMBOL", "name": "attribute_item"},
2333                        {"type": "BLANK"}
2334                    ]},
2335                    {"type": "SYMBOL", "name": "parameter"},
2336                    {"type": "REPEAT", "content": {
2337                        "type": "SEQ",
2338                        "members": [
2339                            {"type": "STRING", "value": ","},
2340                            {"type": "SYMBOL", "name": "parameter"}
2341                        ]
2342                    }}
2343                ]
2344            }"#,
2345        )
2346        .expect("seq");
2347        let symbols = referenced_symbols(&prod);
2348        assert!(symbols.contains(&"attribute_item"));
2349        assert!(symbols.contains(&"parameter"));
2350    }
2351
2352    #[test]
2353    fn literal_strings_collects_choice_members() {
2354        let prod: Production = serde_json::from_str(
2355            r#"{
2356                "type": "CHOICE",
2357                "members": [
2358                    {"type": "STRING", "value": "+"},
2359                    {"type": "STRING", "value": "-"},
2360                    {"type": "STRING", "value": "*"}
2361                ]
2362            }"#,
2363        )
2364        .expect("choice");
2365        let strings = literal_strings(&prod);
2366        assert_eq!(strings, vec!["+", "-", "*"]);
2367    }
2368
2369    /// The ocaml and javascript grammars (tree-sitter ≥ 0.25) emit a
2370    /// `RESERVED` rule kind that an earlier deserialiser rejected
2371    /// with `unknown variant "RESERVED"`. Verify both that the bare
2372    /// variant deserialises and that a `RESERVED`-wrapped grammar is
2373    /// loadable end-to-end via [`Grammar::from_bytes`].
2374    #[test]
2375    fn reserved_variant_deserialises() {
2376        let prod: Production = serde_json::from_str(
2377            r#"{
2378                "type": "RESERVED",
2379                "content": {"type": "SYMBOL", "name": "_lowercase_identifier"},
2380                "context_name": "attribute_id"
2381            }"#,
2382        )
2383        .expect("RESERVED parses");
2384        match prod {
2385            Production::Reserved { content, .. } => match *content {
2386                Production::Symbol { name } => assert_eq!(name, "_lowercase_identifier"),
2387                other => panic!("expected inner SYMBOL, got {other:?}"),
2388            },
2389            other => panic!("expected RESERVED, got {other:?}"),
2390        }
2391    }
2392
2393    #[test]
2394    fn reserved_grammar_loads_end_to_end() {
2395        let bytes = br#"{
2396            "name": "tiny_reserved",
2397            "rules": {
2398                "program": {
2399                    "type": "RESERVED",
2400                    "content": {"type": "SYMBOL", "name": "ident"},
2401                    "context_name": "keywords"
2402                },
2403                "ident": {"type": "PATTERN", "value": "[a-z]+"}
2404            }
2405        }"#;
2406        let g = Grammar::from_bytes("tiny_reserved", bytes).expect("RESERVED-using grammar loads");
2407        assert!(g.rules.contains_key("program"));
2408    }
2409
2410    #[test]
2411    fn reserved_walker_helpers_recurse_into_content() {
2412        // The walker's helpers (first_symbol, has_field_in,
2413        // referenced_symbols, ...) all need to descend through
2414        // RESERVED into its content. If they bail at RESERVED, the
2415        // `pick_choice_with_cursor` heuristic ranks the alt below
2416        // alts that DO recurse, which produces wrong emit output
2417        // even when the deserialiser doesn't crash.
2418        let prod: Production = serde_json::from_str(
2419            r#"{
2420                "type": "RESERVED",
2421                "content": {
2422                    "type": "FIELD",
2423                    "name": "lhs",
2424                    "content": {"type": "SYMBOL", "name": "expr"}
2425                },
2426                "context_name": "ctx"
2427            }"#,
2428        )
2429        .expect("nested RESERVED parses");
2430        assert_eq!(first_symbol(&prod), Some("expr"));
2431        assert!(has_field_in(&prod, &["lhs"]));
2432        let symbols = referenced_symbols(&prod);
2433        assert!(symbols.contains(&"expr"));
2434    }
2435}