@polycode-projects/the-mechanical-code-talker 0.3.0 → 0.4.0

package/src/corpus/conceptnet.mjs

@@ -116,15 +116,27 @@ export function toFacts(assertions, map) {
 
 /** Seed a repo's memory graph (<dir>/.tmct/memory/graph.json) from the
  *  committed slice. Options: limit (cap the facts written — handy for tests
- *  and fast bootstraps), slicePath/mapPath overrides.
+ *  and fast bootstraps), slicePath/mapPath overrides, and `prefer` — an array
+ *  of predicate URIs that STABLE-partitions the facts before the limit is
+ *  applied (facts whose predicate appears earlier in `prefer` come first;
+ *  everything else keeps slice order after them). A capped bootstrap seed
+ *  wants the DEFINITIONAL band ("a cache is a kind of buffer") ahead of the
+ *  location trivia the slice happens to open with; without `prefer` the
+ *  behavior is byte-identical to before.
  *
  *  Idempotent twice over: appendFact's content-hashed ids make a blind
  *  re-append an upsert, and we pre-read the store once to skip triples that
  *  are already there (so re-seeding costs one read, not N rewrites).
  *  Returns { appended, skipped, total }. */
-export async function seedMemory(dir, { limit, slicePath = SLICE_FILE, mapPath = MAP_FILE } = {}) {
+export async function seedMemory(dir, { limit, slicePath = SLICE_FILE, mapPath = MAP_FILE, prefer } = {}) {
   const [assertions, map] = await Promise.all([loadSlice(slicePath), loadMap(mapPath)]);
   let facts = toFacts(assertions, map);
+  if (Array.isArray(prefer) && prefer.length) {
+    const rank = new Map(prefer.map((p, i) => [p, i]));
+    // stable partition: Array.prototype.sort is stable in Node, so equal-rank
+    // facts keep their slice order — deterministic across runs by construction.
+    facts = facts.slice().sort((a, b) => (rank.get(a.predicate) ?? prefer.length) - (rank.get(b.predicate) ?? prefer.length));
+  }
   if (limit !== undefined) facts = facts.slice(0, limit);
 
   // One read up front: what does the store already reify? Keys are built with

package/src/corpus/templates.mjs

@@ -19,9 +19,25 @@ const PKG_ROOT = join(dirname(fileURLToPath(import.meta.url)), "..", "..");
 export const TEMPLATES_FILE = join(PKG_ROOT, "data", "templates", "responses.jsonl");
 export const PHRASEBOOK_FILE = join(PKG_ROOT, "data", "phrasebook", "software-phrases.txt");
 
-const REGISTERS = new Set(["terse", "friendly"]);
+// Registers (Phase 6, PLAN_FORMULAIC_COMPETENCE.md): `terse|friendly` are the
+// conversational bands; `technical` is the C1 / technical-paper band whose
+// templates render item-5 mechanical conclusions (count / comparison /
+// superlative + the provenance we already compute) as advanced prose. A
+// technical template is FORMULAIC COMPETENCE: it renders via:"template", so the
+// dual banding counts it in the PERFORMANCE band only, never the productive one.
+const REGISTERS = new Set(["terse", "friendly", "technical"]);
 const SLOT_RE = /\{([A-Za-z][A-Za-z0-9]*)\}/g;
 
+// Slot-lint for the technical band: a technical template may ONLY fill from the
+// mechanical values tmct actually computes (counts, comparisons, superlatives,
+// scopes, provenance) — no free-text slot can smuggle unattributable prose into
+// the C1 register. Every technical row must also carry a {provenance} fill (the
+// item-5 "+ provenance" contract: an advanced claim always shows its source).
+export const TECHNICAL_SLOTS = Object.freeze(new Set([
+  "subject", "count", "noun", "scope", "comparison", "metric", "unit",
+  "superlative", "provenance",
+]));
+
 /** The slot names a template string requires, in first-appearance order. */
 export function slotsOf(template) {
   const out = [];
@@ -31,6 +47,66 @@ export function slotsOf(template) {
   return out;
 }
 
+// --- Segmentation IR (Phase 7, lever 1) -------------------------------------
+// A rendered answer is ALSO a list of typed spans: [{ type, text }, …] with
+// type ∈ prose | entity | path | number | code | provenance | receipt. The
+// invariant law is byte-exact reconstruction: flatten(segments) === render().
+// Everything except `prose` is PROTECTED — finishing (a later wave) may only
+// transform prose spans, so a grammar rule can never touch a fact.
+//
+// Slot kinds map a template hole to its protected span type. A slot fill is
+// ALWAYS protected (never prose): it is grounded data, not our wording. The
+// specific type is derived from the slot name; unknown slots fall back to the
+// conservative `entity` (protect-when-unsure). Bytes never depend on the type,
+// only on the fill, so the type is metadata layered over an exact split.
+const SLOT_KIND = {
+  count: "number",
+  when: "number",
+  location: "path",
+  commit: "path",
+  provenance: "provenance",
+};
+
+/** The protected span type for a template slot name (default: entity). */
+export function slotKind(name) {
+  return SLOT_KIND[name] || "entity";
+}
+
+/** Reconstruct the flat answer from its segments — a pure, total inverse of
+ *  segmentation: `segments.map(s => s.text).join("")`. flatten(renderSegments(
+ *  id, slots)) === render(id, slots), byte for byte. */
+export function flatten(segments) {
+  let out = "";
+  for (const s of segments) out += s.text;
+  return out;
+}
+
+/** Render template `id` as a SEGMENTED answer: the literal text between slots
+ *  becomes `prose` spans, each slot fill becomes a PROTECTED span typed by
+ *  slotKind(). Validation is identical to render() (unknown id / missing slot
+ *  throw the same messages), so render() is exactly flatten(renderSegments()). */
+export function renderSegments(id, slots = {}, templates = cache) {
+  if (!templates) throw new Error("renderSegments() before loadTemplates() — load the template library first");
+  const row = templates.get(id);
+  if (!row) throw new Error(`unknown template id "${id}"`);
+  const missing = row.slots.filter((s) => slots[s] === undefined || slots[s] === null);
+  if (missing.length) {
+    throw new Error(`template "${id}" missing slot${missing.length > 1 ? "s" : ""}: ${missing.join(", ")}`);
+  }
+  const tpl = row.template;
+  const re = new RegExp(SLOT_RE.source, "g"); // own lastIndex; never touch the shared regex
+  const segments = [];
+  let last = 0;
+  let m;
+  while ((m = re.exec(tpl)) !== null) {
+    if (m.index > last) segments.push({ type: "prose", text: tpl.slice(last, m.index) });
+    segments.push({ type: slotKind(m[1]), text: String(slots[m[1]]) });
+    last = m.index + m[0].length;
+  }
+  if (last < tpl.length) segments.push({ type: "prose", text: tpl.slice(last) });
+  return segments;
+}
+
 let cache = null; // Map<id, row> from the last loadTemplates() — render()'s source
 
 /** Load + validate the response templates. Every line must parse as JSON with
@@ -56,10 +132,21 @@ export async function loadTemplates(path = TEMPLATES_FILE) {
       }
     }
     if (!REGISTERS.has(row.register)) {
-      throw new Error(`${path}:${n + 1}: register must be terse|friendly, got "${row.register}"`);
+      throw new Error(`${path}:${n + 1}: register must be terse|friendly|technical, got "${row.register}"`);
     }
     if (byId.has(row.id)) throw new Error(`${path}:${n + 1}: duplicate template id "${row.id}"`);
-    byId.set(row.id, { ...row, slots: slotsOf(row.template) });
+    const slots = slotsOf(row.template);
+    // Technical-band slot-lint: mechanical-only fills, provenance mandatory.
+    if (row.register === "technical") {
+      const stray = slots.filter((s) => !TECHNICAL_SLOTS.has(s));
+      if (stray.length) {
+        throw new Error(`${path}:${n + 1}: technical template "${row.id}" uses non-mechanical slot${stray.length > 1 ? "s" : ""}: ${stray.join(", ")} (allowed: ${[...TECHNICAL_SLOTS].join(", ")})`);
+      }
+      if (!slots.includes("provenance")) {
+        throw new Error(`${path}:${n + 1}: technical template "${row.id}" must carry a {provenance} fill (an advanced claim always shows its source)`);
+      }
+    }
+    byId.set(row.id, { ...row, slots });
   }
   cache = byId;
   return byId;
@@ -71,13 +158,10 @@ export async function loadTemplates(path = TEMPLATES_FILE) {
  *  `templates` explicitly to bypass the module cache, e.g. in tests). */
 export function render(id, slots = {}, templates = cache) {
   if (!templates) throw new Error("render() before loadTemplates() — load the template library first");
-  const row = templates.get(id);
-  if (!row) throw new Error(`unknown template id "${id}"`);
-  const missing = row.slots.filter((s) => slots[s] === undefined || slots[s] === null);
-  if (missing.length) {
-    throw new Error(`template "${id}" missing slot${missing.length > 1 ? "s" : ""}: ${missing.join(", ")}`);
-  }
-  return row.template.replace(SLOT_RE, (_, name) => String(slots[name]));
+  // render() IS the flattened segmentation, by construction: the byte output is
+  // provably identical to the old `.replace(SLOT_RE, …)` (test/segments.test.mjs
+  // renders every responses.jsonl row both ways and asserts equality).
+  return flatten(renderSegments(id, slots, templates));
 }
 
 /** Load + parse the SE phrase book. Returns:

package/src/finish.mjs

@@ -0,0 +1,443 @@
+// finish.mjs — Phase 7 response finishing: the segmentation IR seam.
+// (PLAN_RESPONSE_FINISHING.md, "The segmentation IR (lever 1)".)
+//
+// The governing principle is fact-invariance BY CONSTRUCTION. An answer is a
+// list of typed spans, [{ type, text }, …], carried alongside the flat string
+// (never replacing it). Every type except `prose` is PROTECTED: entities,
+// paths, numbers, code, provenance and receipts are byte-copied through
+// finishing untouched, and only prose spans are ever handed to a (future)
+// grammar-rule engine. Segmentation makes "turn app/lib/a.mjs into an.mjs"
+// UNREPRESENTABLE — the protected spans are not in the rule engine's input.
+//
+// This module is the FOUNDATION step: pure structure, ZERO behaviour change.
+// It provides:
+//   - the segment type vocabulary + the protected/prose split,
+//   - maskSegments(answer, { graph }) — a conservative masker for the composed
+//     (non-template) path (the templated path segments in corpus/templates.mjs),
+//   - an INVARIANCE CHECKER (the protected-span multiset must survive any
+//     prose-only transform), the property future grammar rules are gated on,
+//   - a NO-OP finish(result, ctx) — the seam a later wave wires into chat.mjs.
+//
+// Byte-exact reconstruction is the whole contract here: flatten(segments) ===
+// answer for every producer, and finish() returns its input byte-for-byte.
+
+import { readFileSync } from "node:fs";
+import { fileURLToPath } from "node:url";
+import { join, dirname } from "node:path";
+import { parse as parseToml } from "smol-toml";
+
+import { flatten } from "./corpus/templates.mjs";
+
+export { flatten };
+
+const GRAMMAR_DIR = dirname(fileURLToPath(import.meta.url));
+/** The data-driven grammar-rule table (Phase 7, lever 2). */
+export const GRAMMAR_RULES_FILE = join(GRAMMAR_DIR, "..", "data", "templates", "grammar-rules.toml");
+
+/** The segment type vocabulary. `prose` is the only unprotected type. */
+export const SEGMENT_TYPES = Object.freeze([
+  "prose", "entity", "path", "number", "code", "provenance", "receipt",
+]);
+
+/** Protected types: byte-copied through finishing, never rule-transformed. */
+export const PROTECTED_TYPES = Object.freeze(
+  new Set(SEGMENT_TYPES.filter((t) => t !== "prose")),
+);
+
+/** Is this span type protected (i.e. anything that is not prose)? */
+export function isProtected(type) {
+  return type !== "prose";
+}
+
+// --- Conservative masker for the composed path ------------------------------
+// The templated path gets segments almost for free (corpus/templates.mjs
+// renderSegments). The composed path (ask engine, plain/conversational turns)
+// hands finishing a hand-built flat string; maskSegments walks it and marks
+// PROTECTED anything matching one of the patterns below, leaving everything
+// else prose. Policy: CONSERVATIVE — when unsure, protect. An un-adopted render
+// site simply presents its whole answer as a single prose span (pass no graph
+// and match nothing → one prose segment), and the invariance checker still
+// guards it. flatten(maskSegments(answer, ctx)) === answer, always.
+
+// Parenthesized receipts: "(traversal: calls edges where object = fnAlpha)" and
+// the repair receipt 'read as "which modules import a.mjs"'.
+const RECEIPT_PAREN_RE = /\((?:traversal|read as)\b[^)]*\)/gi;
+const RECEIPT_READAS_RE = /read as\s+"[^"]*"/gi;
+
+// Provenance / licence tags: "(source: …)", "(licence: …)", bare CC licences
+// ("CC-BY-SA") and the ConceptNet source name.
+const PROVENANCE_PAREN_RE = /\((?:source|licen[sc]e)\b[^)]*\)/gi;
+const LICENCE_CODE_RE = /\bCC-BY(?:-[A-Z]+)*\b/g;
+const CONCEPTNET_RE = /\bConceptNet\b/g;
+
+// Path tokens: anything with a slash-joined segment, or a file extension.
+const PATH_RE = /(?:[\w@.-]+\/)+[\w@.-]+|\b[\w-]+\.(?:mjs|cjs|js|jsx|ts|tsx|json|jsonl|md|txt|py|rb|go|rs|toml|yml|yaml)\b/g;
+
+// Bare numbers (integers or decimals), lowest precedence so a number inside a
+// path/receipt/entity is claimed by that span first.
+const NUMBER_RE = /\b\d+(?:\.\d+)?\b/g;
+
+// Precedence: lower wins when candidate spans overlap.
+const PRECEDENCE = { receipt: 0, provenance: 1, path: 2, entity: 3, number: 4 };
+
+/** Collect the string labels of a loaded graph's individuals (defensive: any
+ *  non-array / missing-label shape yields an empty set). Longest labels first
+ *  so a compound label wins over a substring of it. */
+function graphLabels(graph) {
+  const labels = [];
+  const individuals = graph && Array.isArray(graph.individuals) ? graph.individuals : [];
+  for (const ind of individuals) {
+    const label = ind && ind.label != null ? String(ind.label) : "";
+    if (label) labels.push(label);
+  }
+  labels.sort((a, b) => b.length - a.length);
+  return labels;
+}
+
+const escapeRe = (s) => s.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+
+function pushMatches(re, type, text, out) {
+  re.lastIndex = 0;
+  let m;
+  while ((m = re.exec(text)) !== null) {
+    if (m[0].length === 0) { re.lastIndex += 1; continue; }
+    out.push({ start: m.index, end: m.index + m[0].length, type });
+  }
+}
+
+/** Segment a composed (non-template) flat answer into typed spans, marking as
+ *  PROTECTED any known graph label (entity), path token, bare number, receipt
+ *  tail or provenance/licence tag; everything else is prose. Guarantees exact
+ *  reconstruction: flatten(maskSegments(answer, ctx)) === answer.
+ *
+ *  @param {string} answer   the flat answer string
+ *  @param {{graph?: {individuals?: Array<{label?: any}>}}} [ctx]
+ */
+export function maskSegments(answer, ctx = {}) {
+  const text = String(answer);
+  if (!text) return [];
+  const cand = [];
+
+  pushMatches(RECEIPT_PAREN_RE, "receipt", text, cand);
+  pushMatches(RECEIPT_READAS_RE, "receipt", text, cand);
+  pushMatches(PROVENANCE_PAREN_RE, "provenance", text, cand);
+  pushMatches(LICENCE_CODE_RE, "provenance", text, cand);
+  pushMatches(CONCEPTNET_RE, "provenance", text, cand);
+  pushMatches(PATH_RE, "path", text, cand);
+  pushMatches(NUMBER_RE, "number", text, cand);
+
+  // Known graph labels → entity spans (word-boundary, longest-first).
+  for (const label of graphLabels(ctx && ctx.graph)) {
+    const re = new RegExp(`(?<![\\w/.-])${escapeRe(label)}(?![\\w/.-])`, "g");
+    pushMatches(re, "entity", text, cand);
+  }
+
+  // Greedy non-overlapping selection: earliest start, then precedence, then
+  // longest span. A candidate overlapping an already-accepted span is dropped.
+  cand.sort((a, b) =>
+    a.start - b.start ||
+    PRECEDENCE[a.type] - PRECEDENCE[b.type] ||
+    (b.end - b.start) - (a.end - a.start));
+
+  const chosen = [];
+  let guard = -1;
+  for (const c of cand) {
+    if (c.start < guard) continue;
+    chosen.push(c);
+    guard = c.end;
+  }
+
+  // Emit segments, filling the gaps between protected spans with prose.
+  const segments = [];
+  let last = 0;
+  for (const c of chosen) {
+    if (c.start > last) segments.push({ type: "prose", text: text.slice(last, c.start) });
+    segments.push({ type: c.type, text: text.slice(c.start, c.end) });
+    last = c.end;
+  }
+  if (last < text.length) segments.push({ type: "prose", text: text.slice(last) });
+  return segments;
+}
+
+// --- Invariance checker -----------------------------------------------------
+// The machine proof that no fact moved: the MULTISET of protected spans must be
+// identical before and after any prose-only transform. Future grammar rules are
+// gated on this — a rule that changes a protected span is a bug, not a fix.
+
+/** The protected spans of a segment list, as {type,text} (order-independent). */
+export function protectedSpans(segments) {
+  return segments
+    .filter((s) => isProtected(s.type))
+    .map((s) => ({ type: s.type, text: s.text }));
+}
+
+/** A canonical (sorted) multiset key for the protected spans, for comparison. */
+export function protectedMultiset(segments) {
+  return protectedSpans(segments)
+    .map((s) => `${s.type} ${s.text}`)
+    .sort();
+}
+
+/** Does the protected-span multiset survive a transform (before → after)? */
+export function invariantHolds(before, after) {
+  const a = protectedMultiset(before);
+  const b = protectedMultiset(after);
+  if (a.length !== b.length) return false;
+  for (let i = 0; i < a.length; i += 1) if (a[i] !== b[i]) return false;
+  return true;
+}
+
+/** Assert the invariant, returning `after` on success; throw on violation. */
+export function assertInvariance(before, after) {
+  if (!invariantHolds(before, after)) {
+    throw new Error("fact-invariance violated: the protected-span multiset changed during finishing");
+  }
+  return after;
+}
+
+// --- The grammar-rule engine (lever 2) --------------------------------------
+// applyGrammar transforms ONLY the prose spans of a segment list. It NEVER
+// regexes the flat answer string and NEVER touches a protected span — the
+// invariance checker is treated as NECESSARY-BUT-NOT-SUFFICIENT (a token the
+// masker failed to protect would sit in a prose span, and a corrupting rule that
+// mangled it would pass the multiset check because prose is unchecked). So the
+// only defence is that a rule literally cannot receive a protected span: every
+// handler below filters `type === "prose"` and byte-copies the rest. Each rule's
+// NEUTRAL behaviour is byte-stable; the only byte changes are GENUINE fixes to
+// defects tmct itself generates. Rules are chosen to commute → idempotent.
+
+const escapeRe2 = (s) => String(s).replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+
+/** Preserve the leading-letter case of `orig` on the replacement `want`. */
+function matchCase(orig, want) {
+  return /^[A-Z]/.test(orig) ? want[0].toUpperCase() + want.slice(1) : want;
+}
+
+/** The first alphabetic word token of a string (leading punctuation skipped),
+ *  or "" when there is none (a bare number / symbol → the caller must refuse). */
+function leadingWord(text) {
+  const m = String(text).match(/[A-Za-z][\w-]*/);
+  return m ? m[0] : "";
+}
+
+/** Does `word` begin with a VOWEL SOUND? true → "an", false → "a", null → cannot
+ *  tell (refuse). Spelling-vs-sound exceptions come from the rule's TOML row. */
+function beginsWithVowelSound(word, rule) {
+  const w = String(word).toLowerCase().replace(/^[^a-z]+/, "");
+  if (!w) return null;
+  for (const ex of rule.vowel_sound_consonants || []) if (w.startsWith(ex)) return true;
+  for (const ex of rule.consonant_sound_vowels || []) if (w.startsWith(ex)) return false;
+  const c = w[0];
+  if ("aeiou".includes(c)) return true;
+  if (/[a-z]/.test(c)) return false;
+  return null;
+}
+
+/** Plurality of a span: an explicit boolean `plural` wins; else a NUMBER span is
+ *  singular iff its value is exactly 1 (0 and >1 are plural); else null. */
+function pluralityOf(seg) {
+  if (!seg) return null;
+  if (typeof seg.plural === "boolean") return seg.plural;
+  if (seg.type === "number") {
+    const v = Number(String(seg.text).replace(/,/g, ""));
+    return Number.isFinite(v) ? v !== 1 : null;
+  }
+  return null;
+}
+
+// Rule 1 — article selection (a/an). Reads the following word: in-span when the
+// whole "a word" pair is inside one prose span; across the boundary when the
+// prose ends in "a"/"an" and the next span supplies the word (guard #3: it only
+// fires when it can read the real next token, else it leaves the article alone).
+function ruleArticle(segments, rule) {
+  const out = segments.map((s) => ({ ...s }));
+  for (let i = 0; i < out.length; i += 1) {
+    const seg = out[i];
+    if (seg.type !== "prose") continue;
+    // (a) in-span "a artifact" / "an module" (case-insensitive; case preserved)
+    seg.text = seg.text.replace(/\b(a|an)(\s+)([A-Za-z][\w-]*)/gi, (m, art, sp, word) => {
+      const vowel = beginsWithVowelSound(word, rule);
+      if (vowel === null) return m;
+      return matchCase(art, vowel ? "an" : "a") + sp + word;
+    });
+    // (b) boundary: prose ends "…a " / "…an ", next span carries the word
+    const bm = seg.text.match(/(^|[^\w])(a|an)(\s+)$/i);
+    if (bm) {
+      const next = out[i + 1];
+      const word = next && typeof next.text === "string" ? leadingWord(next.text) : "";
+      const vowel = word ? beginsWithVowelSound(word, rule) : null;
+      if (vowel !== null) {
+        const fixed = matchCase(bm[2], vowel ? "an" : "a");
+        seg.text = seg.text.slice(0, bm.index + bm[1].length) + fixed + bm[3];
+      }
+    }
+  }
+  return out;
+}
+
+// Rule 2 — subject–verb agreement. STRUCTURE-DRIVEN and deliberately narrow:
+// (i) an existential "there is/are/was/were" agrees with the FOLLOWING number
+//     span's value (the number is genuinely the subject there);
+// (ii) any listed copula agrees with a following span that carries an explicit
+//      `plural` flag (a producer that knows its slot's plurality sets it).
+// A bare copula followed by a number that is an OBJECT count ("the file has 3")
+// is NOT existential and carries no flag → left untouched. No surface guessing.
+function ruleAgreement(segments, rule) {
+  const singular = rule.singular || [];
+  const plural = rule.plural || [];
+  const toPlural = {};
+  const toSingular = {};
+  for (let k = 0; k < singular.length; k += 1) { toPlural[singular[k]] = plural[k]; }
+  for (let k = 0; k < plural.length; k += 1) { toSingular[plural[k]] = singular[k]; }
+  const out = segments.map((s) => ({ ...s }));
+  for (let i = 0; i < out.length; i += 1) {
+    const seg = out[i];
+    if (seg.type !== "prose") continue;
+    const m = seg.text.match(/(\bthere\s+)?\b([A-Za-z]+)(\s*)$/i);
+    if (!m) continue;
+    const verb = m[2].toLowerCase();
+    if (!(verb in toPlural) && !(verb in toSingular)) continue;
+    const next = out[i + 1];
+    const existential = Boolean(m[1]) && next && next.type === "number";
+    const flagged = next && typeof next.plural === "boolean";
+    if (!existential && !flagged) continue; // guard: only fire on real structure
+    const isPlural = pluralityOf(next);
+    if (isPlural === null) continue;
+    let want = null;
+    if (isPlural && verb in toPlural && toPlural[verb] !== verb) want = toPlural[verb];
+    else if (!isPlural && verb in toSingular && toSingular[verb] !== verb) want = toSingular[verb];
+    if (!want) continue;
+    seg.text = seg.text.slice(0, m.index) + (m[1] || "") + matchCase(m[2], want) + m[3];
+  }
+  return out;
+}
+
+// Rule 3 — sentence capitalisation. Only when the answer OPENS on a prose span
+// whose first non-space character is a lowercase letter. An answer that opens on
+// a protected span (a path/entity) is left exactly as grounded.
+function ruleCapitalise(segments) {
+  if (!segments.length || segments[0].type !== "prose") return segments;
+  const out = segments.map((s) => ({ ...s }));
+  out[0].text = out[0].text.replace(/^(\s*)([a-z])/, (m, sp, ch) => sp + ch.toUpperCase());
+  return out;
+}
+
+// Rule 4 — list punctuation. A series joined by repeated connective prose spans
+// (default " and ") becomes a comma series with a single terminal conjunction:
+// "a and b and c" → "a, b and c". Operates ONLY on the connective prose spans;
+// the joined (protected) items are byte-copied. A two-item list is untouched.
+function ruleList(segments, rule) {
+  const conn = rule.connective;
+  const sep = rule.separator;
+  const out = segments.map((s) => ({ ...s }));
+  const isConn = (i) => out[i] && out[i].type === "prose" && out[i].text === conn;
+  let i = 0;
+  while (i < out.length) {
+    if (!isConn(i)) { i += 1; continue; }
+    const run = [i];
+    let j = i;
+    while (isConn(j + 2)) { run.push(j + 2); j += 2; }
+    if (run.length >= 2) {
+      for (let k = 0; k < run.length - 1; k += 1) out[run[k]].text = sep; // last stays " and "
+    }
+    i = j + 1;
+  }
+  return out;
+}
+
+// Rule 5 — terminal punctuation. Collapse a trailing run of 2+ sentence stops in
+// the LAST prose span to a single stop ("done.." → "done."). Adds nothing where
+// a fragment/list answer legitimately ends without a stop.
+function ruleTerminal(segments, rule) {
+  const stops = (rule.stops && rule.stops.length ? rule.stops : [".", "!", "?"]).map(escapeRe2).join("");
+  const out = segments.map((s) => ({ ...s }));
+  let idx = -1;
+  for (let i = out.length - 1; i >= 0; i -= 1) if (out[i].type === "prose") { idx = i; break; }
+  if (idx < 0) return out;
+  const re = new RegExp(`([${stops}])(?:\\s*[${stops}])+(\\s*)$`);
+  const m = out[idx].text.match(re);
+  if (m) out[idx].text = out[idx].text.slice(0, m.index) + m[1] + m[2];
+  return out;
+}
+
+const HANDLERS = {
+  article: ruleArticle,
+  agreement: ruleAgreement,
+  capitalise: ruleCapitalise,
+  list: ruleList,
+  terminal: ruleTerminal,
+};
+
+let rulesCache = null;
+
+/** Load + parse the grammar-rule table (the `[[rule]]` array-of-tables). Sync so
+ *  finish() stays synchronous at the turn seam. A bad/absent table is defensive:
+ *  the caller falls back to an empty rule set (finish becomes a strict no-op). */
+export function loadGrammarRules(path = GRAMMAR_RULES_FILE) {
+  const parsed = parseToml(readFileSync(path, "utf8"));
+  return Array.isArray(parsed.rule) ? parsed.rule : [];
+}
+
+/** The cached rule table (parsed once per process). */
+export function grammarRules() {
+  if (rulesCache === null) {
+    try { rulesCache = loadGrammarRules(); } catch { rulesCache = []; }
+  }
+  return rulesCache;
+}
+
+/** Apply the grammar rules to a segment list, in file order, transforming ONLY
+ *  prose spans. Asserts fact-invariance at the end (the protected multiset must
+ *  be identical) and returns the transformed segments. Idempotent by design:
+ *  applyGrammar(applyGrammar(x)) yields the same flattened text as applyGrammar(x). */
+export function applyGrammar(segments, rules = grammarRules()) {
+  let cur = segments;
+  for (const rule of rules) {
+    if (rule && rule.enabled === false) continue;
+    const handler = rule && HANDLERS[rule.kind];
+    if (!handler) continue;
+    cur = handler(cur, rule);
+  }
+  assertInvariance(segments, cur); // necessary-but-not-sufficient; still a hard gate
+  return cur;
+}
+
+// --- The finish() seam ------------------------------------------------------
+// finish() is the LAST transform in a turn:
+//   1. take the result's `segments` (attached by a producer) or mask its
+//      `answer` via maskSegments(result.answer, ctx),
+//   2. map ONLY the prose spans through the grammar-rule engine (applyGrammar),
+//   3. re-flatten, ASSERTING the invariance checker holds (guard #4: this runs
+//      in production, not just in tests — a corruption throws, never ships),
+//   4. rewrite result.answer (and thus logLines); `via` is unchanged.
+//
+// NEUTRAL finishing is BYTE-STABLE: when no rule fires, the flattened answer is
+// byte-identical to the input and finish() returns its argument by REFERENCE, so
+// every byte-exact assertion (test/showcase.test.mjs) stays green untouched. A
+// genuine fix (e.g. "a artifact" → "an artifact") rebuilds the result with the
+// corrected answer. Idempotent by construction: finish(finish(x)) === finish(x).
+//
+// INTENDED chat.mjs SEAM (a sibling/later wave wires this, foreign file): in
+// runTurn, at the `withLast` seam, `result = finish(result, { graph })` so every
+// producer passes through once. Until then finish() is exercised by its unit +
+// golden tests; wiring it changes no fact, only fixes our own generated defects.
+
+/** Finish a turn result: grammar-correct its prose spans, preserving every fact.
+ *  Byte-stable when neutral (returns its argument unchanged); rebuilds only on a
+ *  genuine fix. Throws if finishing would move any protected span (guard #4). */
+export function finish(result, ctx = {}) {
+  if (!result || typeof result.answer !== "string") return result;
+  const before = Array.isArray(result.segments) && result.segments.length
+    ? result.segments
+    : maskSegments(result.answer, ctx);
+  const after = applyGrammar(before, grammarRules());
+  assertInvariance(before, after);
+  const answer = flatten(after);
+  if (answer === result.answer) return result; // neutral → byte-stable, same reference
+  const next = { ...result, answer, segments: after };
+  if (Array.isArray(result.logLines)) {
+    next.logLines = result.logLines.map((l) =>
+      (typeof l === "string" ? l.split(result.answer).join(answer) : l));
+  }
+  return next;
+}

package/src/hash.mjs

@@ -0,0 +1,32 @@
+// hash.mjs — the single home for tmct's content-address hash.
+//
+// FNV-1a 32-bit is deliberately home-grown (see PLAN_DEPENDENCY_STRATEGY.md): it
+// must be synchronous, browser-safe, dependency-free, and — critically —
+// CROSS-VERSION STABLE, because fact ids are content-addressed by it and a fact's
+// id is its identity across the whole memory graph. Every library candidate fails
+// at least one of those; this eight-line function fails none. It lives here, once,
+// so the fact-id contract has exactly one definition.
+//
+// Two historical copies are reconciled here without changing a single output byte:
+//   - src/memory/core.mjs used the hex form for fact ids;
+//   - chatbench/graded.mjs used the integer form as a PRNG seed, with a redundant
+//     mid-loop `>>> 0`. That `>>> 0` was always a no-op: `^` and Math.imul both
+//     apply ToInt32 to their operands, so the 32-bit pattern is invariant between
+//     iterations whether the accumulator is stored signed or unsigned. The final
+//     `h >>> 0` therefore yields the same value either way — proven, not assumed.
+
+/** FNV-1a 32-bit. Returns the unsigned 32-bit integer (0 … 2^32−1). */
+export function fnv1a32(str) {
+  let h = 0x811c9dc5;
+  for (let i = 0; i < str.length; i += 1) {
+    h ^= str.charCodeAt(i);
+    h = Math.imul(h, 0x01000193);
+  }
+  return h >>> 0;
+}
+
+/** FNV-1a 32-bit as a zero-padded 8-char hex string — the stable content-address
+ *  used for fact ids (`fact:<hex>`). Same (s,p,o) → same id → upsert, never a dup. */
+export function fnv1aHex(str) {
+  return fnv1a32(str).toString(16).padStart(8, "0");
+}