comments: the m* sweep — a Russian word in a Macedonian rule, and two truncated sentences

src/languages/macedonian/macedonian.ts

@@ -44,12 +44,12 @@ const mkCountForm = (n: number): number => {
     return m % 10 === 1 && m !== 11 ? 0 : 1;
 };
 
-// #562 symbol normalization — Macedonian. Percent/currency/units/rates carry the count form above
+// symbol normalization — Macedonian. Percent/currency/units/rates carry the count form above
 // (1 процент, 2 проценти; 1 километар, 2 километри). Units are written BOTH ways — Cyrillic кm and Latin
 // km are equally common in the corpus — so both scripts are declared. The rate denominator is "на час"
 // (per hour) but "во секунда" (per second), which the keyed unitPer expresses.
 const SYMBOLS = makeSymbolNormalizer({
-    // #586 `multiply` — the word is this language's OWN, harvested from its existing `×` rule, so nothing new
+    // `multiply` — the word is this language's OWN, harvested from its existing `×` rule, so nothing new
     // is sourced. Declaring it HERE is what makes ASCII `x` read like `×`: `6x6 cm` was reading the `x` as a
     // LETTER NAME, and `NxN` forms outnumber `×` roughly 85 to 20 across the corpora. One word, so `by` is
     // omitted and defaults to it — this language does not split dimension from product.
@@ -173,7 +173,7 @@ const TOKEN = /([а-шА-ШѓѕјљњќџЃЅЈЉЊЌЏѐѝЀЍ]+)|(\d+(?:,\d+)?
 
 class MacedonianPhonemizer implements Phonemizer {
     text(input: string): string {
-        // #562 order: Macedonian rewrites (grouping, ordinals, century/date, clock, ranges, signs) →
+        // order: Macedonian rewrites (grouping, ordinals, century/date, clock, ranges, signs) →
         // INITIALISMS (after abbreviations, so `Д-р` is not spelled DE-ER) → the shared symbol tier last
         // (it needs the number still adjacent to its unit/sign). Roman numerals arrive already converted
         // at the registry seam, so regnal "Лиалофи III" is "3" by the time normalize runs.

src/languages/macedonian/normalize.ts

@@ -41,7 +41,7 @@
  * ORDERING COUPLINGS, each a bug that happened:
  *   · de-grouping FIRST — a period/space is otherwise a token boundary or clause mark.
  *   · multi-dot era markers BEFORE the single-dot year rule, and before the `N г.` expansion.
- *   · the ordinal-suffix rule BEFORE the century/date rules and the range rule.
+ *   · the ordinal-suffix rule BEFORE the century/date rules, which would otherwise claim the digits.
  *   · the range rule BEFORE the clock rule, so `22:00-23:00` becomes "22:00 до 23:00" first.
  *   · personal-initial single capitals BEFORE the initialism pass (which would otherwise see the dot).
  */
@@ -188,7 +188,8 @@ export function normalizeMacedonian(input: string): string {
         return `${a} до ${b}`;
     });
 
-    // 7) CLOCK. The colon was clause punctuation, so `06:30` read as "шест , триесет". Macedonian reads
+    // 6) CLOCK. The colon is otherwise clause punctuation, so `06:30` reads as "шест , триесет".
+    //    Macedonian reads
     //    the time CARDINALLY with "и": 6:30 = "шест и триесет"; :00 drops the minutes (10:00 = "десет").
     //    The trailing "часот"/"ч" (the hour) is kept/expanded — "ч" is the abbreviation for "часот". A
     //    comma after the minutes marks a sports time, not a clock. The range rule ran first, so
@@ -204,12 +205,12 @@ export function normalizeMacedonian(input: string): string {
             return typeof ч === "string" ? `${body} часот` : body;
         },
     );
-    // 7b) `am`/`pm` after a time → претпладне/попладне. Standalone words (the clock already consumed the
+    // 6b) `am`/`pm` after a time → претпладне/попладне. Standalone words (the clock already consumed the
     //     digits), matched on word boundaries so "Сами" is untouched.
     s = s.replace(/\bam\b/giu, "претпладне");
     s = s.replace(/\bpm\b/giu, "попладне");
 
-    // 8) THE ORDINAL SUFFIX — the big one. `(\d+)[- ]?` + the written suffix (the last letters of the
+    // 7) THE ORDINAL SUFFIX, the largest class in this file. `(\d+)[- ]?` + the written suffix (the last letters of the
     //    spoken ordinal). The `-те` suffix is ambiguous between a count ("the N") and a decade: a 4-digit
     //    number followed by "години" is a decade, otherwise it is "the N".
     s = s.replace(
@@ -241,9 +242,9 @@ export function normalizeMacedonian(input: string): string {
         },
     );
 
-    // 9) CENTURY — `N век` → ordinal + век (10 век → десетти век), including the compound list
+    // 8) CENTURY — `N век` → ordinal + век (10 век → десетти век), including the compound list
     //    `10 и 11 век`. Also the one Germanic remnant `8. век` (with a dot). The suffix forms (17-ти век,
-    //    18-тиот век) were already claimed by step 8.
+    //    18-тиот век) are already claimed by step 7.
     s = s.replace(/(\d+)\s+и\s+(\d+)\s*век(?![\p{L}\p{M}])/gu, (_m, a: string, b: string) => {
         const oa = mkOrdinal(Number(a)), ob = mkOrdinal(Number(b));
         return oa !== undefined && ob !== undefined ? `${oa} и ${ob} век` : _m;
@@ -257,7 +258,7 @@ export function normalizeMacedonian(input: string): string {
         return o === undefined ? m0 : `${o} век`;
     });
 
-    // 10) DATES — `N месяц` → ordinal + месяц (на 6 октомври → на шести октомври), including a date range
+    // 9) DATES — `N <month>` → ordinal + month (на 6 октомври → на шести октомври), including a date range
     //    `24 август - 5 септември` and the Germanic-dot form `4. јули 1776`. The day must be 1–31.
     const monthAlt = [...MONTHS].join("|");
     s = s.replace(
@@ -277,9 +278,7 @@ export function normalizeMacedonian(input: string): string {
         },
     );
 
-    // 11) (the range rule lives at step 5, before the clock.)
-
-    // 11) REGNAL ORDINALS. All three corpus Romans (Лиалофи III, Елизабета II, Луј XVI) are regnal proper
+    // 10) REGNAL ORDINALS. All three corpus Romans (Лиалофи III, Елизабета II, Луј XVI) are regnal proper
     //     names, and the shared Roman pass has already rewritten them to CARDINAL digits before this engine
     //     runs (Macedonian has no ROMAN_POLICIES entry). The digit after a capitalized NAME is read as an
     //     ordinal: Лиалофи 3 → Лиалофи трети, Луј 16 → Луј шеснаесетти. Guarded by the same ≤39 monarch
@@ -297,7 +296,7 @@ export function normalizeMacedonian(input: string): string {
         return `${name} ${name.endsWith("а") ? femIndef(o) : o}`;
     });
 
-    // 12) RATE UNITS the shared tier cannot compose. `милји/час` (miles/hour — милји is a full word, not an
+    // 11) RATE UNITS the shared tier cannot compose. `милји/час` (miles/hour — милји is a full word, not an
     //     abbreviation), the Latin `mph`/`kph`, `Mbit/s` (megabits per second). The Cyrillic squared units
     //     `мм2`/`км2` are also local: the tier's exponent lookbehind `(?<=[a-zA-Z])` is ASCII-only and the
     //     corpus writes `3136 мм2`.
@@ -308,15 +307,15 @@ export function normalizeMacedonian(input: string): string {
     s = s.replace(/(\d+)\s*мм\s*[²2](?!\d)/gu, "$1 квадратни милиметри");
     s = s.replace(/(\d+)\s*км\s*[²2](?!\d)/gu, "$1 квадратни километри");
 
-    // 13) DEGREES — `90°F`, `35° W`, and a bare `N°`. The corpus's own spelled-out form is
+    // 12) DEGREES — `90°F`, `35° W`, and a bare `N°`. The corpus's own spelled-out form is
     //     "30 степени целзиусови", so °C/°F use the "по" construction; `° W`/`° E` are coordinates.
     s = s.replace(/(\d+)\s*°\s*C(?![\p{L}\p{M}])/gu, "$1 степени по Целзиус");
     s = s.replace(/(\d+)\s*°\s*F(?![\p{L}\p{M}])/gu, "$1 степени по Фаренхаjт");
     s = s.replace(/(\d+)\s*°\s*W(?![\p{L}\p{M}])/gu, "$1 степени запад");
     s = s.replace(/(\d+)\s*°\s*E(?![\p{L}\p{M}])/gu, "$1 степени исток");
     s = s.replace(/(\d+)\s*°/gu, "$1 степени");
 
-    // 14) SIGNS. `+30` (the corpus's "над +30 степени") reads "плус". Minus, ×, ÷, =, <, > and the
+    // 13) SIGNS. `+30` (the corpus's "над +30 степени") reads "плус". Minus, ×, ÷, =, <, > and the
     //     ampersand are the handoff's sign classes — none occurs in the corpus but a dropped sign is
     //     inaudible, so each is read.
     s = s.replace(/(^|[\s(])[-−]\s?(?=\d)/gu, "$1минус ");
@@ -334,12 +333,12 @@ export function normalizeMacedonian(input: string): string {
     s = s.replace(/\s*>\s*/gu, " поголемо од ");
     s = s.replace(/\s*[&＆]\s*/gu, " и ");
 
-    // 15) pH → "пе ха" (letter names) and `Ghz` → "гигахерци" — the corpus's lowercase-tech tokens that
+    // 14) pH → "пе ха" (letter names) and `Ghz` → "гигахерци" — the corpus's lowercase-tech tokens that
     //     would otherwise read as consonant clusters or Latin foreign.
     s = s.replace(/(?<![\p{L}\p{M}])pH(?![\p{L}\p{M}])/gu, "пе ха");
     s = s.replace(/(?<![\p{L}\p{M}])Ghz(?![\p{L}\p{M}])/giu, "гигахерци");
 
-    // 15b) FRACTIONS — the corpus's "29¾ инчи на 24½ инчи" and "5 мм (1/5 инчи)". ¾/½ after a whole read
+    // 14b) FRACTIONS — the corpus's "29¾ инчи на 24½ инчи" and "5 мм (1/5 инчи)". ¾/½ after a whole read
     //     "и три четвртини" / "и половина"; a unit fraction reads "една петтина"-shaped (numeral-ordinand
     //     stem + -тина). The vulgar-fraction glyphs are not in any clause-punctuation map, so they were
     //     being dropped outright.
@@ -354,7 +353,7 @@ export function normalizeMacedonian(input: string): string {
         return `една ${suffix}`;
     });
 
-    // 16) PERSONAL-INITIAL single capitals — `Н. Вејн` (the shared LONE_INITIAL handles `В. Буш`, which
+    // 15) PERSONAL-INITIAL single capitals — `Н. Вејн` (the shared LONE_INITIAL handles `В. Буш`, which
     //     sits between capitalized words; this one follows a comma in "НАСА, Н. Вејн"). And a lone Latin
     //     capital standing as a LETTER between words — `буквата V` (the letter vee) reads "ве" not the
     //     English [viː]; `H во pH` (after pH → "пе ха") reads "ха".

src/languages/madurese/madurese.ts

@@ -137,10 +137,10 @@ export function phonemizeWord(word: string): string {
 const TOKEN = new RegExp(`(${hostWordRun(["Latin"], "'")})|(\\d+)|([.!?…,;:])`, "giu");
 
 /**
- * This language's OWN inventory — the TOKEN word class as it stood before the widening above, lifted
- * verbatim, so nothing about the orthography is invented here. A token this REJECTS carries a letter the
- * language does not use, i.e. a foreign name. See core/hostWord.ts: this is the INVENTORY question, and it
- * is no longer also deciding where the script boundary falls.
+ * This language's OWN inventory. ⚠ TWO DIFFERENT QUESTIONS, KEPT APART: the TOKEN class above decides where
+ * the SCRIPT boundary falls (routing), while this one decides whether the g2p has rules for these letters. A
+ * token this class REJECTS carries a letter the language does not use — i.e. a foreign name. See
+ * core/hostWord.ts.
  */
 const NATIVE_CLASS = "[a-zâèéòḍṭ']";
 const nat = makeNativiser(NATIVE_CLASS, "iu");

src/languages/magahi/magahi.ts

@@ -1,24 +1,25 @@
 /**
- * Native Magahi / मगही (mag) text phonemizer — canonical IPA. Indo-Aryan (Bihari, Magadhan),
- * Devanagari. Reuses the Hindi/Bhojpuri Devanagari engine (makeNativeHindi — schwa deletion, weight stress) with a
- * Magahi data file. Magahi shares the Bihari core with Bhojpuri — NO phonemic vowel length, single sibilant श/ष→[s],
- * ण/ञ→[n] — but differs in its documented GLIDE HARDENING (Vinod Kumar 2026, A Comparative Phonological Study of
- * Bihari Languages, §6.2): word-initial व→[b] (वंश→bans) and य→[d͡ʒ] (यन्त्र→jantar), where Bhojpuri preserves the
- * glides (व→w, य→j). Single comparative-source delta on the grammar-anchored Bhojpuri base → 🔷. Was a mag→bho
- * alias until the reference revealed the delta.
+ * Native Magahi / मगही (mag) text phonemizer — canonical IPA. Indo-Aryan (Bihari, Magadhan), Devanagari.
+ * Reuses the Hindi/Bhojpuri Devanagari engine (makeNativeHindi — schwa deletion, weight stress) with a Magahi
+ * data file.
+ *
+ * Magahi shares the Bihari core with Bhojpuri — NO phonemic vowel length, single sibilant श/ष→[s], ण/ञ→[n] —
+ * but differs in its documented GLIDE HARDENING (Vinod Kumar 2026, *A Comparative Phonological Study of Bihari
+ * Languages*, §6.2): word-initial व→[b] (वंश→bans) and य→[d͡ʒ] (यन्त्र→jantar), where Bhojpuri preserves the
+ * glides (व→w, य→j). ⚠ That delta rests on a SINGLE comparative source over the grammar-anchored Bhojpuri base.
  */
 /**
- * NORMALIZER WORDS: the Hindi defaults are RETAINED and four are confirmed for Magahi, one of them by a
- * DEFINITION rather than a usage. Evidence from mag.wikipedia via `tools/normalization/attest.ts`
- * (`tools/corpus/attest/mag.jsonc`); sentences judged Magahi by हे / हल / हलै / आउ / -के:
+ * NORMALIZER WORDS. This engine inherits Hindi's, and four are confirmed for Magahi — sentences judged
+ * Magahi by हे / हल / हलै / आउ / -के:
+ *
+ *   प्रतिशत    a DEFINITIONAL citation, which outranks any usage example: "प्रतिशतके अर्थ हे प्रति सौ या
+ *             प्रति सैकड़ा (% = 1/100)" ties the word directly to the SIGN this layer is reading.
+ *   बजे       "रातके १२ बजे होलै हल"
+ *   मिनट      "प्रति मिनट क्रान्ति, सङ्क्षिप्त आर/मिनट या आरपिएम हे"
+ *   ईसा पूर्व  "ईसा पूर्व ३०४ से ईसा पूर्व २३२ … हल"
  *
- *   प्रतिशत  ✓✓ the Magahi wiki's own article DEFINES it: "प्रतिशतके अर्थ हे प्रति सौ या प्रति सैकड़ा
- *               (% = 1/100)" — a definitional citation tying the word to the SIGN, which is stronger than
- *               any usage example. Also in use: "साक्षरता दर ७५ प्रतिशत … हलै आउ".
- *   बजे      ✓ "रातके १२ बजे होलै हल"
- *   मिनट     ✓ "प्रति मिनट क्रान्ति, सङ्क्षिप्त आर/मिनट या आरपिएम हे"
- *   ईसा पूर्व ✓ "ईसा पूर्व ३०४ से ईसा पूर्व २३२ … हल"
- *   बजकर      NOT ATTESTED (0 hits) — Hindi's clock connective stands, unconfirmed.
+ * बजकर, Hindi's clock connective, is NOT attested for Magahi. It stands unconfirmed rather than being
+ * replaced — an unsourced substitute is worse than an inherited word.
  */
 import { makeNativeHindi, type HindiDef, type ForeignPhonemizer } from "../hindi/hindi.ts";
 import { loadManifest } from "../../core/loadManifest.ts";

src/languages/maithili/maithili.ts

@@ -1,26 +1,25 @@
 /**
- * Native Maithili / मैथिली (mai) text phonemizer — canonical IPA. Eastern Indo-Aryan (Bihari
- * group), Devanagari. Reuses the Hindi Devanagari engine (makeNativeHindi — schwa deletion, weight stress,
- * numbers) with a Maithili data file whose divergences from Hindi are: SHORT e/o (ए→e, ओ→o), the diphthongs
- * ऐ→[əɪ] / औ→[əu], and inherent /ə/. Maithili's signature — a cluster schwa that Hindi DELETES instead reduces to
- * an ULTRASHORT [ᵊ] (इसपात→ɪsᵊpaːt) — is a narrow phonetic detail folded (ᵊ~∅) against the referee.
+ * Native Maithili / मैथिली (mai) text phonemizer — canonical IPA. Eastern Indo-Aryan (Bihari group),
+ * Devanagari. Reuses the Hindi Devanagari engine (makeNativeHindi — schwa deletion, weight stress, numbers)
+ * with a Maithili data file. The divergences from Hindi: SHORT e/o (ए→e, ओ→o), the diphthongs ऐ→[əɪ] /
+ * औ→[əu], and inherent /ə/.
  *
- * 🔷 SINGLE-SOURCE: the only referee is wikipron mai_deva narrow (167 human pairs) — small, so this is a
- * single-source-verified bring-up, not a confident convergence. Distinct from the ⛔ Bhojpuri/Awadhi stubs, which
- * have NO referee at all.
+ * Maithili's signature — a cluster schwa that Hindi DELETES is instead reduced to an ULTRASHORT [ᵊ]
+ * (इसपात→ɪsᵊpaːt) — is a narrow phonetic detail, folded (ᵊ~∅) rather than contrasted.
+ *
+ * ⚠ SINGLE-SOURCE bring-up: the only referee is a small human set (167 pairs), so this is verified rather
+ * than convergent.
  */
 /**
- * NORMALIZER WORDS: Hindi's CLOCK words are confirmed for Maithili; its percent word is NOT, and is
- * retained unconfirmed. Evidence from mai.wikipedia via `tools/normalization/attest.ts`
- * (`tools/corpus/attest/mai.jsonc`). This language is why the reading has to be done per sentence:
+ * NORMALIZER WORDS. This engine inherits Hindi's. The CLOCK words (बजकर, मिनट) are confirmed for Maithili;
+ * the percent word प्रतिशत is NOT, and is retained unconfirmed — an unsourced substitute is worse than an
+ * inherited word.
  *
- *   बजकर + मिनट ✓ "प्राण प्रतिष्ठाक मुहूर्त १२ बजकर २९ मिनट ८ सेकेण्ड सँ … रहल जे"
- *                  — प्रतिष्ठाक / सेकेण्ड सँ / रहल जे are Maithili. TWO of the four बजकर hits are Maithili;
- *                    the other two are the same passage in HINDI (का मुहूर्त, रहा, मिलाकर), quoted verbatim.
- *   मिनट     ✓ "5 घंटा 30 मिनट जोड़ला सं प्राप्त होइत अछि"           — होइत अछि is Maithili
- *   प्रतिशत   ✗ its single hit is a NEPALI passage ("९७.९ प्रतिशत भन्दा अधिक … लागेकाछन्"), so it says
- *              nothing about Maithili. Hindi's default stands, unconfirmed — not replaced, because an
- *              unsourced substitute is worse than an inherited word (this issue's own standing rule).
+ * ⚠ ATTESTATION MUST BE JUDGED PER SENTENCE, NOT PER HIT, and this language is why. Devanagari wikis quote
+ * each other verbatim, so a search for a Maithili word returns Hindi and Nepali passages under a Maithili
+ * title: half the बजकर hits here are one Hindi passage quoted whole (का मुहूर्त, रहा, मिलाकर), and the
+ * only प्रतिशत hit is Nepali (भन्दा अधिक, लागेकाछन्). A hit counts only when the surrounding sentence
+ * carries Maithili morphology — होइत अछि, रहल जे, सेकेण्ड सँ.
  */
 import { makeNativeHindi, type HindiDef, type ForeignPhonemizer } from "../hindi/hindi.ts";
 import { loadManifest } from "../../core/loadManifest.ts";