parseman 0.46.0 → 0.47.0

package/src/spec/model.ts

@@ -0,0 +1,636 @@
+/**
+ * Grammar spec model — a small, notation-agnostic tree that both the EBNF
+ * emitter and the railroad-diagram emitter consume.
+ *
+ * The model is produced by walking the SAME `_def` structure the interpreter
+ * and macro compiler consume, so a generated spec cannot drift from what
+ * actually parses. See `docs/proposals/grammar-spec-generation.md`.
+ */
+import type { Combinator, ParserDef } from '../types.ts'
+import type { BalancedMarked } from '../combinators/scanTo.ts'
+import { RULE_ORDER } from '../combinators/parser.ts'
+import { recoverComposedRules } from '../compiler/linker.ts'
+
+// ---------------------------------------------------------------------------
+// Spec node tree
+// ---------------------------------------------------------------------------
+
+/** A node in the grammar spec tree — one per syntactic construct. */
+export type SpecNode =
+  | { kind: 'seq'; items: SpecNode[] }
+  | { kind: 'choice'; items: SpecNode[] }
+  /** `item*`, or `item{0,max}` when `many(…, { max })` bounded it. */
+  | { kind: 'star'; item: SpecNode; max?: number }
+  /** `item+`, or `item{min,max}` when `many(…, { min, max })` bounded it. `min`
+   *  defaults to 1 and is always >= 1 (a min-0 repeat is a `star`). */
+  | { kind: 'plus'; item: SpecNode; min?: number; max?: number }
+  | { kind: 'opt'; item: SpecNode }
+  /**
+   * Separated repetition. `min`/`max` count ITEMS, not separators: `min: 0` is
+   * `(item (sep item)*)?` — NULLABLE — and any `min >= 1` requires that many
+   * items, so it is not. `trailing` mirrors `sepBy`'s option of the same name.
+   */
+  | { kind: 'sepBy'; item: SpecNode; sep: SpecNode; min: number; max?: number; trailing?: 'allow' }
+  /** Reference to a named production (non-terminal). */
+  | { kind: 'ref'; name: string }
+  /**
+   * A terminal. `literal` terminals are exact strings (rendered quoted);
+   * non-literal terminals are patterns/prose (rendered as-is, e.g. a regex
+   * or a caller-supplied display name).
+   */
+  | { kind: 'terminal'; text: string; literal: boolean }
+  /** Negative lookahead (`not`). Rendered as an annotation, not consumed input. */
+  | { kind: 'not'; item: SpecNode }
+  /** Positive lookahead (`ahead`). Zero-width, like `not`. */
+  | { kind: 'peek'; item: SpecNode }
+  /** An out-of-band annotation (guards, error-recovery, unknowns). */
+  | { kind: 'annotation'; text: string }
+  /** Matches nothing (elided trivia / semantic-only wrappers). */
+  | { kind: 'empty' }
+
+/** One named production: `name ::= expr`. */
+export type Production = {
+  name: string
+  expr: SpecNode
+  /** True when the rule is trivia (whitespace/comment). Elided by default. */
+  trivia: boolean
+}
+
+export type SpecModel = {
+  productions: Production[]
+}
+
+export type SpecOptions = {
+  /**
+   * How to order the emitted productions when neither `order` nor `root` is given:
+   *   - `'source'` (default) — the order the rules were **declared** in the
+   *     `rules()` factory. Predictable, includes every rule, and leads with the
+   *     entry rule (authors write it first). Internal-but-referenced helper rules
+   *     follow, in first-reference order.
+   *   - `'reachable'` — BFS from the entry rule: the first-declared rule leads,
+   *     then each rule appears the first time it is referenced; any rules not
+   *     reachable from the entry are appended in declaration order.
+   * Ignored when `order` or `root` is set.
+   */
+  sort?: 'source' | 'reachable'
+  /**
+   * Explicit rule order (and subset). When given, only these rules are emitted,
+   * in this order, plus any rules they reach. Overrides `sort`.
+   */
+  order?: string[]
+  /**
+   * Start rule(s). When given, only these and the rules they reach are emitted
+   * (a pruned, reachability-ordered spec). Overrides `sort`.
+   */
+  root?: string | string[]
+  /** Include trivia rules (whitespace/comment) in the output. Default: false. */
+  includeTrivia?: boolean
+  /**
+   * Display names for terminals, keyed by rule name. When a rule whose body is a
+   * single terminal (regex/literal/keywords) has an entry here, that whole rule
+   * renders as the given terminal name instead of expanding — e.g.
+   * `{ Ident: 'identifier' }`.
+   */
+  terminals?: Record<string, string>
+  /**
+   * Best-effort rendering of a regex terminal to a readable form. Return
+   * `undefined` to fall back to the default `/source/` rendering.
+   */
+  regexDisplay?: (source: string, flags: string) => string | undefined
+}
+
+export type GrammarInput =
+  | Record<string, Combinator<unknown>>
+  | Combinator<unknown>
+
+// ---------------------------------------------------------------------------
+// Walk
+// ---------------------------------------------------------------------------
+
+type Tagged = Combinator<unknown> & { _ruleName?: string }
+
+function ruleNameOf(c: Combinator<unknown>): string | undefined {
+  return (c as Tagged)._ruleName
+}
+
+/** Resolve a `lazy`/`ref` combinator one hop to its defined body, if resolvable. */
+function resolveLazy(c: Combinator<unknown>): Combinator<unknown> | undefined {
+  const def = c._def
+  if (def.tag !== 'lazy') return undefined
+  try {
+    return def.thunk() as Combinator<unknown>
+  } catch {
+    return undefined
+  }
+}
+
+function defaultRegexDisplay(source: string): string {
+  return `/${source}/`
+}
+
+// ---------------------------------------------------------------------------
+// Authored terminals behind a regex
+// ---------------------------------------------------------------------------
+//
+// A spec reader's vocabulary is the LANGUAGE's: MDN and the CSS specs draw
+// `@import`, not `/@import(?![-_a-zA-Z0-9\u0080-\uFFFF])/`. The trailing
+// lookahead is a word-boundary assertion — an implementation detail of how a
+// keyword is recognised, not part of the language — and printing it drowns the
+// grammar in emitter noise. It also broke the diagrams as a complexity metric:
+// a rule made of keywords scored worse than a genuinely tangled one.
+//
+// So when a `regex()` terminal IS a fixed string (optionally guarded by a word
+// boundary, optionally a set of fixed strings), render the string(s). This is
+// not simplification — nothing is collapsed, hidden, or elided; the terminal is
+// shown as the text it matches, which is strictly MORE information than its
+// compiled pattern. Anything with real regex structure still prints raw.
+
+/**
+ * Start index of the top-level group that closes at the very end of `src`, or
+ * `-1`. Character classes and escapes are skipped so `[)]` and `\)` don't lie.
+ */
+function trailingGroupStart(src: string): number {
+  let depth = 0
+  let open = -1
+  let lastStart = -1
+  let lastEnd = -1
+  for (let i = 0; i < src.length; i++) {
+    const ch = src[i]!
+    if (ch === '\\') { i++; continue }
+    if (ch === '[') {
+      // Skip the class body.
+      for (i++; i < src.length; i++) {
+        if (src[i] === '\\') i++
+        else if (src[i] === ']') break
+      }
+      continue
+    }
+    if (ch === '(') { if (depth === 0) open = i; depth++ }
+    else if (ch === ')') { depth--; if (depth === 0) { lastStart = open; lastEnd = i } }
+  }
+  return lastEnd === src.length - 1 ? lastStart : -1
+}
+
+/** `(?![…])` — a negative lookahead over a single character class. */
+const BOUNDARY_GUARD = /^\(\?!\[(?:[^\\\]]|\\[\s\S])*\]\)$/
+
+/** Split on top-level `|`, skipping escapes, classes and groups. */
+function topLevelAlternatives(src: string): string[] {
+  const parts: string[] = []
+  let depth = 0
+  let start = 0
+  for (let i = 0; i < src.length; i++) {
+    const ch = src[i]!
+    if (ch === '\\') { i++; continue }
+    if (ch === '[') {
+      for (i++; i < src.length; i++) {
+        if (src[i] === '\\') i++
+        else if (src[i] === ']') break
+      }
+      continue
+    }
+    if (ch === '(') depth++
+    else if (ch === ')') depth--
+    else if (ch === '|' && depth === 0) { parts.push(src.slice(start, i)); start = i + 1 }
+  }
+  parts.push(src.slice(start))
+  return parts
+}
+
+/** Regex metacharacters an escape may legitimately spell as a literal. */
+const ESCAPABLE = '\\^$.|?*+()[]{}/-'
+
+/**
+ * Decode one alternative to the exact string it matches, or `undefined` when it
+ * is not a fixed string. Deliberately strict: any quantifier, class, group,
+ * anchor, class shorthand or control escape disqualifies it, because rendering
+ * those as text would state something the grammar does not say.
+ */
+function fixedString(alt: string): string | undefined {
+  if (alt.length === 0) return undefined
+  let out = ''
+  for (let i = 0; i < alt.length; i++) {
+    const ch = alt[i]!
+    if (ch === '\\') {
+      const next = alt[i + 1]
+      if (next === undefined) return undefined
+      if (next === 'u' || next === 'x') {
+        const hex = next === 'u'
+          ? (alt[i + 2] === '{' ? /^\\u\{([0-9a-fA-F]+)\}/.exec(alt.slice(i)) : /^\\u([0-9a-fA-F]{4})/.exec(alt.slice(i)))
+          : /^\\x([0-9a-fA-F]{2})/.exec(alt.slice(i))
+        if (hex === null) return undefined
+        const cp = parseInt(hex[1]!, 16)
+        // Control and non-printable code points have no honest text form.
+        if (cp < 0x20 || cp === 0x7f) return undefined
+        out += String.fromCodePoint(cp)
+        i += hex[0]!.length - 1
+        continue
+      }
+      if (!ESCAPABLE.includes(next)) return undefined
+      out += next
+      i++
+      continue
+    }
+    // Unescaped metacharacter → real structure, not a fixed string.
+    if ('^$.|?*+()[]{}'.includes(ch)) return undefined
+    // Literal control characters likewise have no honest text form.
+    if (ch.codePointAt(0)! < 0x20) return undefined
+    out += ch
+  }
+  return out.length > 0 ? out : undefined
+}
+
+/**
+ * The fixed string(s) a regex matches — the terminal its author actually wrote —
+ * or `undefined` when the pattern has real structure. Handles the three shapes
+ * that spell a keyword: a bare literal, a literal guarded by a trailing word
+ * boundary (`word()` / `keywords()` and every hand-written copy of them), and a
+ * `\b…\b`-anchored word.
+ */
+export function fixedStringsOfRegex(source: string): string[] | undefined {
+  let src = source
+  // `\b` anchors assert a boundary, exactly like the `(?![…])` guard.
+  if (src.startsWith('\\b')) src = src.slice(2)
+  if (src.endsWith('\\b') && !src.endsWith('\\\\b')) src = src.slice(0, -2)
+
+  const g = trailingGroupStart(src)
+  if (g >= 0 && BOUNDARY_GUARD.test(src.slice(g))) src = src.slice(0, g)
+
+  // `keywords()` compiles to `(?:a|b|c)` plus the guard; unwrap that hull.
+  const outer = trailingGroupStart(src)
+  if (outer === 0 && src.startsWith('(?:')) src = src.slice(3, -1)
+
+  const alts = topLevelAlternatives(src)
+  const out: string[] = []
+  for (const alt of alts) {
+    const fixed = fixedString(alt)
+    if (fixed === undefined) return undefined
+    out.push(fixed)
+  }
+  return out.length > 0 ? out : undefined
+}
+
+class Builder {
+  private seen = new Set<string>()
+  /** Combinators on the CURRENT walk path — the anonymous-cycle guard. */
+  private walking = new Set<Combinator<unknown>>()
+  private pending: Array<{ name: string; comb: Combinator<unknown> }> = []
+  private pendingNames = new Set<string>()
+  private record: Record<string, Combinator<unknown>>
+  private opts: SpecOptions
+  productions: Production[] = []
+
+  constructor(record: Record<string, Combinator<unknown>>, opts: SpecOptions) {
+    this.record = record
+    this.opts = opts
+  }
+
+  enqueue(name: string, comb: Combinator<unknown>): void {
+    if (this.seen.has(name)) return
+    if (this.pendingNames.has(name)) return
+    this.pendingNames.add(name)
+    this.pending.push({ name, comb })
+  }
+
+  /**
+   * Seed and drain in phases. Each phase seeds its names, then BFS-drains the
+   * queue (following references) before the next phase seeds. This lets callers
+   * express "entry first, then its reachable closure, then any leftovers" by
+   * splitting the seed across phases; a single phase with every rule yields plain
+   * declaration order (references, already queued, keep their declared slot).
+   */
+  run(phases: string[][]): SpecModel {
+    for (const phase of phases) {
+      for (const name of phase) {
+        const comb = this.record[name]
+        if (comb) this.enqueue(name, comb)
+      }
+      this.drain()
+    }
+    return { productions: this.productions }
+  }
+
+  private drain(): void {
+    while (this.pending.length > 0) {
+      const { name, comb } = this.pending.shift()!
+      this.pendingNames.delete(name)
+      if (this.seen.has(name)) continue
+      this.seen.add(name)
+      const expr = this.ruleBody(comb, name)
+      this.productions.push({ name, expr, trivia: comb._meta.isTrivia === true })
+    }
+  }
+
+  /** Walk a rule's top-level combinator, transparently unwrapping its own self-ref. */
+  private ruleBody(comb: Combinator<unknown>, name: string): SpecNode {
+    // A rule that IS a single terminal can be pinned to a caller display name.
+    const pinned = this.opts.terminals?.[name]
+    if (pinned !== undefined) return { kind: 'terminal', text: pinned, literal: false }
+
+    let c = comb
+    // Unwrap the rule's own placeholder ref(s) so we expand its body, not a
+    // self-reference. Both the placeholder and the resolved body carry the same
+    // `_ruleName`, so guard against cycles.
+    const guard = new Set<Combinator<unknown>>()
+    while (c._def.tag === 'lazy' && ruleNameOf(c) === name && !guard.has(c)) {
+      guard.add(c)
+      const next = resolveLazy(c)
+      if (!next) break
+      c = next
+    }
+    return this.walk(c, true)
+  }
+
+  /** Walk any combinator. `isRuleRoot` suppresses treating it as a self-reference. */
+  private walk(c: Combinator<unknown>, isRuleRoot = false): SpecNode {
+    // A reference to a named rule → non-terminal (unless this IS that rule's root).
+    const rn = ruleNameOf(c)
+    if (rn !== undefined && !isRuleRoot) {
+      // Register the target so it gets its own production, even if it wasn't in
+      // the returned record (internal-but-referenced helper rules).
+      if (!this.record[rn]) this.record[rn] = c
+      this.enqueue(rn, c)
+      return { kind: 'ref', name: rn }
+    }
+
+    // A balanced() is a delimiter scan, so render the delimiters and mark the
+    // interior opaque — see `BalancedMarked`. Checked before the cycle guard
+    // because it is the honest rendering, not merely a way to terminate.
+    const bal = (c as BalancedMarked)._balanced
+    if (bal !== undefined) {
+      return {
+        kind: 'seq',
+        items: [
+          { kind: 'terminal', text: bal.open, literal: true },
+          { kind: 'annotation', text: 'balanced …' },
+          { kind: 'terminal', text: bal.close, literal: true },
+        ],
+      }
+    }
+
+    // CYCLE CUT ON OBJECT IDENTITY.
+    //
+    // The walk used to cut only at `_ruleName`, which assumes every cycle passes
+    // through a named rule. A combinator whose interior refers back to ITSELF
+    // through an anonymous `ref()` — what `balanced()` builds, and what any caller
+    // of the public `ref()` can build — closes a cycle carrying no name, and the
+    // walk went around it until the process died (RangeError at the default stack,
+    // SIGSEGV at a raised one: a true cycle, not deep-but-finite recursion).
+    //
+    // `walking` is a PATH set, added on entry and removed on exit — not a global
+    // visited set. A combinator shared by two sibling positions is real structure
+    // that must be drawn twice; only a combinator that contains ITSELF is a cycle.
+    if (this.walking.has(c)) return { kind: 'annotation', text: '(recursive)' }
+    this.walking.add(c)
+    try {
+      return this.walkDef(c._def, c)
+    } finally {
+      this.walking.delete(c)
+    }
+  }
+
+  private walkDef(def: ParserDef, self: Combinator<unknown>): SpecNode {
+    switch (def.tag) {
+      case 'literal':
+        return { kind: 'terminal', text: def.value, literal: true }
+
+      case 'regex': {
+        // A caller-supplied display name always wins — it is the most specific
+        // statement of intent available.
+        const shown = this.opts.regexDisplay?.(def.source, def.flags)
+        if (shown !== undefined) return { kind: 'terminal', text: shown, literal: false }
+        const fixed = fixedStringsOfRegex(def.source)
+        if (fixed !== undefined) return literalTerminals(fixed)
+        return { kind: 'terminal', text: defaultRegexDisplay(def.source), literal: false }
+      }
+
+      case 'keywords':
+        return literalTerminals(def.words)
+
+      case 'sequence': {
+        const items = def.parsers.map(p => this.walk(p)).filter(nonEmpty)
+        return flattenSeq(items)
+      }
+
+      case 'choice': {
+        const items = def.parsers.map(p => this.walk(p)).filter(nonEmpty)
+        return flattenChoice(items)
+      }
+
+      case 'dispatch': {
+        const selector = this.walk(def.selector)
+        const tails = [
+          ...def.cases.map(entry => this.walk(entry.parser)),
+          ...(def.matchers ? def.matchers.map(entry => this.walk(entry.parser)) : []),
+          ...(def.otherwise ? [this.walk(def.otherwise)] : []),
+        ]
+        return flattenSeq([selector, flattenChoice(tails)])
+      }
+
+      // `max`/`min` are carried through so a BOUNDED repeat does not render as an
+      // unbounded one — the spec is generated to be read as the truth about what
+      // parses, and `many(x, { min: 3, max: 8 })` is not `x+`.
+      case 'many':
+        return { kind: 'star', item: this.walk(def.parser), ...(def.max === undefined ? {} : { max: def.max }) }
+      case 'oneOrMore':
+        return {
+          kind: 'plus', item: this.walk(def.parser),
+          ...(def.min === 1 ? {} : { min: def.min }),
+          ...(def.max === undefined ? {} : { max: def.max }),
+        }
+      case 'optional':
+        return { kind: 'opt', item: this.walk(def.parser) }
+      case 'attempt':
+        return this.walk(def.parser)
+      case 'sepBy':
+        return {
+          kind: 'sepBy', item: this.walk(def.parser), sep: this.walk(def.separator), min: def.min,
+          ...(def.max === undefined ? {} : { max: def.max }),
+          ...(def.trailing === undefined ? {} : { trailing: def.trailing }),
+        }
+
+      case 'not':
+        return { kind: 'not', item: this.walk(def.parser) }
+      case 'peek':
+        return { kind: 'peek', item: this.walk(def.parser) }
+      case 'routed':
+        // The routed token is spelled by the dispatch selector, so it contributes
+        // nothing here; a fallback IS spelled in place, so render it.
+        return def.fallback === undefined ? { kind: 'empty' } : this.walk(def.fallback)
+
+      // Transparent semantic wrappers — render the inner syntax.
+      case 'transform':
+      case 'token':
+      case 'leaf':
+      case 'label':
+      case 'field':
+      case 'expect':
+      case 'withCtx':
+      case 'node':
+        return this.walk((def as { parser: Combinator<unknown> }).parser)
+
+      case 'grammar':
+        return this.walk(def.parser)
+
+      // Error recovery — render the intended sub-parser, drop the sentinel.
+      case 'recover':
+        return this.walk(def.parser)
+      case 'scanTo':
+        return { kind: 'annotation', text: '…' }
+
+      // Trivia — elided by default; included as an annotation when asked.
+      case 'trivia':
+        return this.opts.includeTrivia ? this.walk(def.parser) : { kind: 'empty' }
+
+      case 'guard':
+        return { kind: 'empty' }
+
+      // Zero-width, but — unlike a state guard — it is a statement ABOUT THE SOURCE
+      // TEXT, so a grammar spec that elided it would document a different language.
+      case 'adjacency':
+        return { kind: 'annotation', text: def.polarity === 'adjacent' ? '(adjacent)' : def.kinds ? `(separated: ${def.kinds.join('|')})` : '(separated)' }
+
+      case 'lazy': {
+        const inner = resolveLazy(self)
+        return inner ? this.walk(inner) : { kind: 'annotation', text: '?' }
+      }
+
+      case 'unknown':
+      default:
+        return { kind: 'annotation', text: '?' }
+    }
+  }
+}
+
+/**
+ * One or more fixed strings as terminal node(s). A ONE-word set is a terminal,
+ * not a one-arm alternation: `word('@import')` used to render `("@import")`,
+ * and drew a branch box in the diagram, purely because a keyword set is
+ * internally a set. That paren and that box are emitter artifacts — there is no
+ * alternative to choose between.
+ */
+function literalTerminals(words: readonly string[]): SpecNode {
+  const items = words.map(w => ({ kind: 'terminal', text: w, literal: true }) as SpecNode)
+  return flattenChoice(items)
+}
+
+function nonEmpty(n: SpecNode): boolean {
+  return n.kind !== 'empty'
+}
+
+function flattenSeq(items: SpecNode[]): SpecNode {
+  const flat: SpecNode[] = []
+  for (const it of items) {
+    if (it.kind === 'seq') flat.push(...it.items)
+    else flat.push(it)
+  }
+  if (flat.length === 0) return { kind: 'empty' }
+  if (flat.length === 1) return flat[0]!
+  return { kind: 'seq', items: flat }
+}
+
+function flattenChoice(items: SpecNode[]): SpecNode {
+  const flat: SpecNode[] = []
+  for (const it of items) {
+    if (it.kind === 'choice') flat.push(...it.items)
+    else flat.push(it)
+  }
+  if (flat.length === 0) return { kind: 'empty' }
+  if (flat.length === 1) return flat[0]!
+  return { kind: 'choice', items: flat }
+}
+
+function toRecord(grammar: GrammarInput): Record<string, Combinator<unknown>> {
+  // A single combinator → a one-rule grammar keyed by its rule name or "start".
+  if (isCombinator(grammar)) {
+    const name = ruleNameOf(grammar) ?? 'start'
+    return { [name]: grammar }
+  }
+  // A `compose()` result is a map of FUSED rule FUNCTIONS, not combinators — walking
+  // it read `_def` off a function and threw a bare `TypeError: Cannot read properties
+  // of undefined (reading 'tag')`. The combinator graph is recoverable from the
+  // carried IR, so recover it rather than fail. (Same recovery the gating analysis
+  // uses — see `recoverComposedRules`.)
+  const recovered = recoverComposedRules(grammar as Record<string, unknown>)
+  if (recovered !== undefined) {
+    if (recovered.opaque.length > 0) {
+      // Fail LOUDLY and specifically. A partial spec is worse than none: it renders a
+      // grammar that looks complete while silently omitting whole artifacts.
+      const named = recovered.opaque.map(o => `"${o.ns}"${o.ruleNames.length > 0 ? ` (${o.ruleNames.length} rule(s))` : ''}`).join(', ')
+      throw new TypeError(
+        `parseman spec: this composed grammar includes opaque precompiled artifact(s) ${named} that carry compiled rule `
+        + 'functions rather than re-lowerable IR. Their rules cannot be rendered, and emitting a spec without them would '
+        + 'silently understate the grammar. Recompile the contributing grammar(s) so they carry IR.',
+      )
+    }
+    return Object.fromEntries(recovered.rules)
+  }
+  return grammar
+}
+
+function isCombinator(x: unknown): x is Combinator<unknown> {
+  return (
+    typeof x === 'object' &&
+    x !== null &&
+    typeof (x as { parse?: unknown }).parse === 'function' &&
+    '_def' in (x as object)
+  )
+}
+
+/** Declaration order recorded by `rules()`; falls back to the record's key order. */
+function declarationOrder(record: Record<string, Combinator<unknown>>): string[] {
+  const declared = (record as Record<string, unknown>)[RULE_ORDER]
+  if (Array.isArray(declared)) return declared.filter(k => k in record)
+  return Object.keys(record)
+}
+
+/**
+ * Build the notation-agnostic spec model from a `rules()` grammar (or a single
+ * combinator). Following every referenced named rule keeps the closure complete;
+ * ordering is controlled by `order` / `root` / `sort` (see {@link SpecOptions}).
+ */
+export function buildSpecModel(grammar: GrammarInput, options: SpecOptions = {}): SpecModel {
+  const record = toRecord(grammar)
+  const decl = declarationOrder(record)
+
+  // Validate rule names in `order` / `root` up front. Left unchecked, an unknown
+  // name (or a stray string like `order: 'source'` where a `string[]` is meant)
+  // seeds a phase that reaches nothing and silently yields an EMPTY model — a
+  // confusing footgun. Fail loudly with the offending names and the known rules.
+  const known = new Set(decl)
+  const asList = (v: string | string[]): string[] => (Array.isArray(v) ? v : [v])
+  const requireKnown = (names: string[], opt: string): void => {
+    const bad = names.filter(n => !known.has(n))
+    if (bad.length > 0) {
+      throw new Error(
+        `buildSpecModel: unknown rule name(s) in \`${opt}\`: ${bad.map(n => JSON.stringify(n)).join(', ')}. ` +
+        `Known rules: ${decl.map(n => JSON.stringify(n)).join(', ')}.`,
+      )
+    }
+  }
+
+  // Seed phases (see Builder.run). Priority: explicit `order` > `root` (pruned) >
+  // `sort`. Default `sort` = 'source' (declaration order, every rule).
+  let phases: string[][]
+  if (options.order !== undefined) {
+    const order = asList(options.order as string | string[])
+    requireKnown(order, 'order')
+    phases = [order]
+  } else if (options.root !== undefined) {
+    const root = asList(options.root)
+    requireKnown(root, 'root')
+    phases = [root]
+  } else if (options.sort === 'reachable') {
+    // Entry (first-declared) leads; BFS discovers the rest in first-reference
+    // order; unreachable rules trail in declaration order.
+    phases = decl.length > 0 ? [[decl[0]!], decl.slice(1)] : [decl]
+  } else {
+    phases = [decl]
+  }
+
+  const model = new Builder({ ...record }, options).run(phases)
+  if (!options.includeTrivia) {
+    model.productions = model.productions.filter(p => !p.trivia)
+  }
+  return model
+}