@capsiynau/intelligence 0.5.0 → 0.5.1

package/CHANGELOG.md

@@ -6,6 +6,46 @@ All notable changes to `@capsiynau/intelligence` go here. Format follows
 
 ## [Unreleased]
 
+## [0.5.1] - 2026-07-25
+
+**Welsh review of the 0.5.0 seed lexicon. Four of the twelve entries were
+wrong, two of them badly enough to emit ungrammatical Welsh.** Patch rather
+than minor because no consumer resolves 0.5.x yet (`notes-app` and `postio`
+are both on `^0.4.0`, and caret on a 0.x range permits patch only), so this
+corrects a defect nobody has installed.
+
+### Fixed
+
+- **`efo`/`gyda` and `isio`/`moyn` are now detect-only.** Both are genuine
+  regional pairs, but substituting them changes the grammar of the words
+  around them, and this module is mutation-blind and syntax-blind by design.
+  `gyda` takes the aspirate mutation and `efo` takes none, so the bare swap
+  gave `gyda cath` where the correct form is `gyda chath`. `moyn` takes a
+  linking `yn` and `isio` does not, so the bare swap gave `Dw i moyn`, which
+  is not grammatical. `checkDialect` still reports both; `applyDialect` now
+  skips them. Emitting broken Welsh is worse than leaving a term alone and
+  telling a human.
+- **Removed `pres`/`arian` (money) and `dŵad`/`dod` (to come).** These are not
+  regional pairs. `arian` and `dod` are the universal standard forms, used in
+  the north as readily as the south; `pres` and `dŵad` are northern
+  colloquialisms. Filing them as regional meant telling a northern writer that
+  their standard Welsh was southern. Register belongs to the prompt fragment,
+  not this table. The reason is recorded in `EXCLUDED_TERMS` so the next person
+  does not re-add them.
+
+### Added
+
+- `applySafe` on every `checkDialect` finding and every `listDialectTerms`
+  row, plus `applyNote` on detect-only entries. A caller can now tell the
+  difference between "this was corrected" and "this needs a human", rather
+  than assuming a clean `applyDialect` pass resolved everything.
+
+### Note
+
+The lexicon has now had a machine-assisted review against house style and
+known Welsh error patterns. That is **not** a native sign-off.
+`DIALECT_TERMS_STATUS.reviewed` stays `false`.
+
 ## [0.5.0] - 2026-07-25
 
 Additive. No existing import path changes shape, so 0.4.x consumers can

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@capsiynau/intelligence",
-  "version": "0.5.0",
+  "version": "0.5.1",
   "description": "Bilingual (Welsh + English) intelligence layer: text normalisation, mutation handling, digraph splitting, spellcheck primitives, and glossary/correction utilities. Pure functions, runs in Node or browser.",
   "type": "module",
   "license": "MIT",

package/src/dialect/index.js

@@ -27,6 +27,13 @@
  *     `welsh/mutations.js`. Callers normalise first or accept the miss.
  *   - Touch pronouns or copula forms. See EXCLUDED_TERMS below; the false
  *     positives are worse than the misses.
+ *   - Rewrite a term whose swap would change the grammar around it. Those
+ *     entries carry `applySafe: false`: checkDialect REPORTS them, applyDialect
+ *     SKIPS them. "efo" -> "gyda" is the clear case, since gyda takes the
+ *     aspirate mutation and efo takes none, so the bare swap is ungrammatical.
+ *   - Carry colloquial-versus-standard pairs. "arian" and "dod" are standard
+ *     everywhere; "pres" and "dŵad" are northern colloquialisms, not southern
+ *     opposites. That is register, and register belongs to the prompt fragment.
  *   - Infer a region from text. The caller declares it, per the house rule
  *     that source language and variety are never auto-detected.
  *
@@ -62,18 +69,38 @@ const WELSH_LETTERS = "A-Za-zÂÊÎÔÛŴŶâêîôûŵŷÁÉÍÓÚáéíóúÀ
  */
 export const DIALECT_TERMS = [
   { id: 'now',         north: 'rŵan',    south: 'nawr',     neutral: 'nawr',    gloss: 'now' },
-  { id: 'with',        north: 'efo',     south: 'gyda',     neutral: 'gyda',    gloss: 'with' },
   { id: 'milk',        north: 'llefrith', south: 'llaeth',  neutral: 'llaeth',  gloss: 'milk' },
-  { id: 'want',        north: 'isio',    south: 'moyn',     neutral: 'eisiau',  gloss: 'to want' },
   { id: 'out',         north: 'allan',   south: 'mas',      neutral: 'allan',   gloss: 'out' },
-  { id: 'money',       north: 'pres',    south: 'arian',    neutral: 'arian',   gloss: 'money' },
-  { id: 'to-come',     north: 'dŵad',    south: 'dod',      neutral: 'dod',     gloss: 'to come' },
   { id: 'boy',         north: 'hogyn',   south: 'crwt',     neutral: 'bachgen', gloss: 'boy' },
   { id: 'girl',        north: 'hogan',   south: 'croten',   neutral: 'merch',   gloss: 'girl' },
   { id: 'cuppa',       north: 'panad',   south: 'dishgled', neutral: 'paned',   gloss: 'a cuppa' },
+
   // Both forms regionally marked, no neutral standard.
   { id: 'grandfather', north: 'taid',    south: 'tad-cu',   neutral: null,      gloss: 'grandfather' },
   { id: 'grandmother', north: 'nain',    south: 'mam-gu',   neutral: null,      gloss: 'grandmother' },
+
+  // ── Detect-only: real regional pairs, but SUBSTITUTING them changes the
+  // grammar of the surrounding words, and this module is mutation-blind and
+  // syntax-blind by design. checkDialect reports them (always safe, and it is
+  // exactly what a reviewer wants to see); applyDialect refuses to rewrite
+  // them, because a swap that produces ungrammatical Welsh is worse than no
+  // swap at all.
+  {
+    id: 'with', north: 'efo', south: 'gyda', neutral: 'gyda', gloss: 'with',
+    applySafe: false,
+    applyNote:
+      '"gyda" triggers the aspirate mutation and "efo" triggers none, so a ' +
+      'bare swap yields "gyda cath" where the correct form is "gyda chath". ' +
+      'Fixing it needs the mutation machinery this module does not have.',
+  },
+  {
+    id: 'want', north: 'isio', south: 'moyn', neutral: 'eisiau', gloss: 'to want',
+    applySafe: false,
+    applyNote:
+      'The three forms take different syntax: "Dw i isio" and "Dw i eisiau" ' +
+      'take no linking "yn", but "moyn" does ("Dw i\'n moyn"). A bare swap ' +
+      'yields "Dw i moyn", which is not grammatical.',
+  },
 ]
 
 /**
@@ -82,6 +109,17 @@ export const DIALECT_TERMS = [
  * they are absent is not self-evident.
  */
 export const EXCLUDED_TERMS = [
+  {
+    id: 'colloquial-not-regional',
+    reason:
+      'Pairs like "pres"/"arian" (money) and "dŵad"/"dod" (to come) look ' +
+      'regional but are not: "arian" and "dod" are the universal standard ' +
+      'forms, used in the north as readily as the south, and "pres"/"dŵad" ' +
+      'are northern COLLOQUIAL forms. That is a register distinction, and ' +
+      'register is the prompt fragment\'s job, not this table\'s. Filing them ' +
+      'here would have told a northern writer their standard Welsh was ' +
+      'southern. Removed 25.07.2026 after review.',
+  },
   {
     id: 'pronoun-o-e',
     reason:
@@ -171,7 +209,9 @@ function activeEntries(region) {
       seen.add(key)
       alternatives.push(form)
     }
-    if (alternatives.length) out.push({ entry, preferred, alternatives })
+    if (alternatives.length) {
+      out.push({ entry, preferred, alternatives, applySafe: entry.applySafe !== false })
+    }
   }
   return out
 }
@@ -184,11 +224,15 @@ function activeEntries(region) {
  * @returns {Array<{id:string, gloss:string, preferred:string, alternatives:string[]}>}
  */
 export function listDialectTerms(region) {
-  return activeEntries(region).map(({ entry, preferred, alternatives }) => ({
+  return activeEntries(region).map(({ entry, preferred, alternatives, applySafe }) => ({
     id: entry.id,
     gloss: entry.gloss,
     preferred,
     alternatives,
+    applySafe,
+    // Present only on detect-only entries, so a UI can explain to the user
+    // why this one is flagged but never auto-corrected.
+    ...(entry.applyNote ? { applyNote: entry.applyNote } : {}),
   }))
 }
 
@@ -205,12 +249,20 @@ export function listDialectTerms(region) {
 export function checkDialect(text, region) {
   if (!text) return []
   const findings = []
-  for (const { entry, preferred, alternatives } of activeEntries(region)) {
+  for (const { entry, preferred, alternatives, applySafe } of activeEntries(region)) {
     for (const alternative of alternatives) {
       const matcher = matcherFor(alternative)
       let hit
       while ((hit = matcher.exec(text)) !== null) {
-        findings.push({ id: entry.id, found: hit[1], expected: preferred, index: hit.index })
+        findings.push({
+          id: entry.id,
+          found: hit[1],
+          expected: preferred,
+          index: hit.index,
+          // false means applyDialect will NOT fix this one: the caller must
+          // hand it to a human rather than assume a rewrite resolved it.
+          applySafe,
+        })
       }
     }
   }
@@ -224,6 +276,12 @@ export function checkDialect(text, region) {
  * Substitution is whole-term only and mutation-blind, so it is safe to run on
  * mixed Welsh/English text: no entry in the table is an English word.
  *
+ * Entries marked `applySafe: false` are SKIPPED here and reported by
+ * `checkDialect` instead. Those are pairs whose substitution would change the
+ * grammar of neighbouring words (mutation or linking "yn"), which this module
+ * has no machinery to get right. Emitting ungrammatical Welsh is worse than
+ * leaving the term alone and telling a human about it.
+ *
  * @param {string} text
  * @param {'north'|'south'|'neutral'|'broadcast'|'formal'} region
  * @returns {string}
@@ -231,7 +289,8 @@ export function checkDialect(text, region) {
 export function applyDialect(text, region) {
   if (!text) return ''
   let out = text
-  for (const { preferred, alternatives } of activeEntries(region)) {
+  for (const { preferred, alternatives, applySafe } of activeEntries(region)) {
+    if (!applySafe) continue
     for (const alternative of alternatives) {
       out = out.replace(matcherFor(alternative), (m) => matchCase(m, preferred))
     }

package/types/dialect/index.d.ts

@@ -34,6 +34,12 @@ export function checkDialect(text: string, region: "north" | "south" | "neutral"
  * Substitution is whole-term only and mutation-blind, so it is safe to run on
  * mixed Welsh/English text: no entry in the table is an English word.
  *
+ * Entries marked `applySafe: false` are SKIPPED here and reported by
+ * `checkDialect` instead. Those are pairs whose substitution would change the
+ * grammar of neighbouring words (mutation or linking "yn"), which this module
+ * has no machinery to get right. Emitting ungrammatical Welsh is worse than
+ * leaving the term alone and telling a human about it.
+ *
  * @param {string} text
  * @param {'north'|'south'|'neutral'|'broadcast'|'formal'} region
  * @returns {string}
@@ -51,13 +57,23 @@ export namespace DIALECT_TERMS_STATUS {
  * standard, so a neutral target leaves the text alone rather than inventing
  * one. Saying "no neutral form exists" is more useful than picking a side.
  */
-export const DIALECT_TERMS: {
+export const DIALECT_TERMS: ({
     id: string;
     north: string;
     south: string;
     neutral: string;
     gloss: string;
-}[];
+    applySafe?: undefined;
+    applyNote?: undefined;
+} | {
+    id: string;
+    north: string;
+    south: string;
+    neutral: string;
+    gloss: string;
+    applySafe: boolean;
+    applyNote: string;
+})[];
 /**
  * Deliberately NOT in the table, with the reason. Kept in code because the
  * next person to extend this will reach for exactly these, and the reason