fix(gloss): Caesar audit fixes — impetus/pagus/barba/leuci/itis core + truncation + Gai citations (M-023l)

Whole-BG-I LLM audit (1968 lemmas) flagged 383 rows; browser triage showed
most are candidate-lemma collisions the popup ranking suppresses. Fixed the
6 confirmed gloss defects:

- core overrides + golden rows: impetus 'an attack, assault, onset, charge'
  (was rare sense[3] 'a fit, paroxysm'), pagus 'a district, canton, commune'
  (was 'the country'), barba 'beard' (WORDS narrative clause), leucus/leuci
  'the Leuci, a people of Gallia Belgica near the Lingones' (was truncated
  'whence, perh'), itis 'you (pl.) go, proceed' (was L&S Tuscan ghost 'Day').
- bestClause clause-split guard: don't split ',\s+the X' after an appositive
  hedge (perh./whence/i.e./viz./scil.) — fixes the leuci truncation mechanism.
  Verified only leucus/leuci affected (removing 'the' from the lookahead
  changed 698 rows and was reverted).
- Gai citation-tail strip ([.,;:]\s*Gai(?:us)?[\s\S]*$, case-sensitive):
  removes 'Gai 2, 190' / 'Gai Inst. 1, 115 sq' from 13 juristic rows
  (certioro, condictio, usurecipio...). NOT a CIT entry (leaves bare-number
  remnants) and not case-insensitive (ate 'gain'/'gait' in 7 glosses).

npm test green (22 unit + 81 IPA + 2075 gloss + 348 census); 9 popup e2e;
Caesar BG-I 1928/1929 glossed.

Author: hellpanderrr

wiktionary_pron/docs/ISSUES.md

@@ -1233,3 +1233,38 @@ src/core/Tokenization.ts.
 **Status: FIXED.** New popup e2e test (Germanis → "The Germans", not "full
 sister"); 9 popup + macronizer e2e + 22 unit + 81 IPA + 2069 golden + 348
 census, all green.
+
+## M-023l — Caesar gloss audit: fix confirmed defects (2026-08-13) ✅
+
+**Symptom.** A whole-BG-I gloss audit (1,968 lemmas via gemini-3.5-flash-lite,
+batch 40, concurrency 16) flagged 383 rows (19.5%). Triage with a live browser
+probe showed the bulk were candidate-lemma collisions the popup ranking
+suppresses (sine→Sinis, eos→Eos, more→morus…) and auditor false-positives
+(explorator "a prying person" is L&S's own sense[0]). The real defects:
+
+- **impetus** → "a fit, paroxysm" (rare sense[3]); L&S sense[0] is "an attack,
+  assault, onset". Shows on every impetus/impetum in Caesar.
+- **pagus** → "the country"; Caesar means "a district, canton".
+- **barba** → "The ancient Romans allowed the beard to grow long" (WORDS
+  narrative first clause); should be "beard".
+- **leucus/leuci** → truncated "…whence, perh" (clause-split at ", the mod.
+  Liège" orphaned the dangling hedge); should be a clean ethnic gloss.
+- **itis** → "Day" (L&S Tuscan ghost); should be "you (pl.) go, proceed" (2pl
+  of eo).
+- **certioro** → trailing ", Gai 2" (Gaius cited WITHOUT a period, so the
+  citation strips miss it).
+
+**Fix.**
+- Core overrides for impetus, pagus, barba, leucus, leuci, itis (each with a
+  golden row; golden 2069→2075).
+- `bestClause` clause-split guard: don't split ",\s+the X" after an appositive
+  hedge (perh./whence/i.e./viz./scil.) — fixes the leuci truncation mechanism
+  (was: split at ", the mod. Liège"). Verified: only leucus/leuci affected.
+- Gai citation-tail strip (`[.,;:]\s*Gai(?:us)?[\s\S]*$`, case-sensitive):
+  removes "Gai 2, 190" / "Gai Inst. 1, 115 sq" from 13 juristic-gloss rows.
+  NOT a CIT-list entry — "Gai" in CIT left a bare-number remnant ("…Ulpian) 2")
+  and case-insensitive matching ate "gain"/"gait" in 7 real glosses.
+
+**Status: FIXED.** Artifact 34,413 lemmas / 450 KB / L&S 89.8%; golden 2075,
+census 348, Caesar BG-I 1928/1929 (only the ide proper-noun false-positive);
+9 popup e2e green.

wiktionary_pron/docs/LESSONS.md

@@ -760,3 +760,33 @@ Catilinam I, Vergil Aeneid I, Ovid Metamorphoses I.
   or BEHIND the upstream `latin-macronizer-wasm` repo. Before editing, diff the two
   to learn which is the source of truth — do NOT `npm run build` + sync blindly, or
   you'll clobber whichever side carries the newer fixes.
+
+## A whole-corpus LLM audit is mostly noise for a good pipeline — but the real finds are worth it (2026-08-13)
+
+- **Symptom.** Full Caesar BG-I gloss audit (1,968 lemmas, gemini-3.5-flash-lite
+  batch audit) → 383 flags (19.5%).
+- **Triage lesson.** ~300 of the 383 were **candidate-lemma collisions** — the
+  wordlist maps a common form to a proper-noun/ghost lemma (`sine`→`Sinis`,
+  `eos`→`Eos`, `more`→`morus`, `saepe`→`saepes`). The audit flags the *candidate
+  universe* (my prep script collected ALL lemmas a form maps to), but the popup
+  shows the *ranked reading*. A live browser probe of 16 high-frequency words
+  proved the ranking suppresses every wrong candidate — **verify candidate-class
+  flags in the browser, never trust lemma-universe counts.**
+- **The real finds were gloss-quality defects invisible to lemma-level checks**:
+  `impetus`→"a fit, paroxysm" (extractor picked rare sense[3] over sense[0]),
+  `pagus`→"the country" (context-primary mismatch), `barba`→a WORDS narrative
+  clause, `leucus`→truncated "whence, perh", `itis`→"Day" (L&S Tuscan ghost),
+  `certioro`→trailing ", Gai 2".
+- **Two build traps hit while fixing the truncation:**
+  1. Removing `the\s` from the `bestClause` split lookahead changed **698 rows**
+     (68 lost glosses) — the ", the X" split is load-bearing. The narrow fix is a
+     **lookbehind before the comma** (`(?<!perh\.)(?<!whence)(?<!i\.e\.)(?<!viz\.)
+     (?<!scil\.),`) — and a lookbehind placed AFTER `,\s+` checks the wrong
+     position (past the whitespace), silently doing nothing.
+  2. Adding "Gai" to the CIT author list stripped ", Gai" but left a bare
+     number ("…Ulpian) 2"); and a case-INSENSITIVE `Gai(?:us)?` ate "gain" and
+     "gait" in 7 real glosses (merito2 "To earn, gain"→"To earn"). The working
+     form is a case-sensitive citation-tail strip: `[.,;:]\s*Gai(?:us)?[\s\S]*$`.
+- **Lesson.** When a build change "fixes" a rare case, always diff the FULL
+  artifact before/after (the 698-row and 7-row blow-ups only showed up there),
+  and keep new citation strippers case-sensitive with a sentence-boundary anchor.

wiktionary_pron/macronizer/glosses.tsv.gz

@@ -266,6 +266,11 @@ function cleanOne(t) {
   // the "de or" inside English words — "asiDE OR away" → declino "to turn asi",
   // "disquietUDE OR confusion" → interturbo "To produce disquietu".
   t = t.replace(/[.,;:]\s+\bde\s+Or\.?[\s\S]*$/i,"").trim();
+  // "Gai 2, 190", "Gai Inst. 1, 115 sq" — Gaius is cited WITHOUT the period (so
+  // the 3+-letter CIT author list and CIT_FULL miss it), leaving "Gai Inst" or a
+  // bare "Gai 2" in the gloss (certioro "...used by Ulpian), Gai 2"; astipulator
+  // "...stipulation, Gai Inst"). Strip the whole ", Gai…" citation to the end.
+  t = t.replace(/[.,;:]\s*Gai(?:us)?[\s\S]*$/,"").trim();
   t = t.replace(CIT,"");
   t = t.replace(/\([^)]*\)\s*$/,"").trim();
   t = t.replace(/\([^)]*$/,"").trim();   // unclosed trailing paren (clause-split broke "(syn.: ...)" across clauses)
@@ -566,7 +571,7 @@ function isGlossyOpen(raw) {
   return first.length >= 3 && (EN_WORDS.has(first) || BIO_WORDS.has(first) || /(?:ous|ful|able|ible|ive|ish|less|like|some|ed|ing)$/.test(first));
 }
 function bestClause(s, pos) {
-  const clauses = s.split(/[;:]/).flatMap(c => c.split(/,\s+(?=to\s|a\s|an\s|the\s|of\s|in\s|by\s|for\s|from\s|with\s|who\s|which\s|one\s|that\s|much\b|few\b|little\b|many\b|great\b|upper\b|lower\b|former\b|latter\b)/))
+  const clauses = s.split(/[;:]/).flatMap(c => c.split(/(?<!perh\.)(?<!whence)(?<!i\.e\.)(?<!viz\.)(?<!scil\.),\s+(?=to\s|a\s|an\s|the\s|of\s|in\s|by\s|for\s|from\s|with\s|who\s|which\s|one\s|that\s|much\b|few\b|little\b|many\b|great\b|upper\b|lower\b|former\b|latter\b)/))
     // definitions that follow a closing etymology bracket: "Engl. else], another, other"
     .flatMap(c => c.split(/\]\s*,?\s+(?=a\b|an\b|the\b|one\b|who\b|which\b|what\b|to\b|another\b|other\b)/))
     // L&S chains subordinate notes with an em-dash ("a probe, Cels. 7, 17.—Hence,

wiktionary_pron/utils/build_glosses.cjs

@@ -226,6 +226,7 @@
   "bacchylides": "a Greek lyric poet, contemporary of Pindar",
   "badia": "a kind of cloak, capote",
   "ballista": "a catapult, engine for hurling missiles",
+  "barba": "beard",
   "barcas": "a surname of Hamilcar",
   "basilica": "a basilica, public hall",
   "beatus": "blessed, happy, fortunate; wealthy",
@@ -878,6 +879,7 @@
   "immortalis": "immortal, undying, everlasting",
   "immunitas": "immunity, exemption",
   "impertior": "to impart, bestow, give a share of",
+  "impetus": "an attack, assault, onset, charge",
   "impositio": null,
   "imprimis": "especially, particularly, in the first place",
   "improfessus": "unavowed, undeclared, unprofessed",
@@ -943,6 +945,7 @@
   "ischias": null,
   "iste": "that of yours; that",
   "ita": "so, thus, in this manner; yes",
+  "itis": "you (pl.) go, proceed (2nd pl. of eo)",
   "iudex": "a judge",
   "iudicium": "a judgment, decision, trial; discernment",
   "iugum": "a yoke; a ridge, mountain-range; a pair (of yoked animals)",
@@ -1004,9 +1007,11 @@
   "leporinus": "of a hare, leporine",
   "lerne": "Lerna, a marshy region of Argos",
   "lethaea": "the wife of Olenus, turned to stone for her pride",
+  "leuci": "the Leuci, a people of Gallia Belgica near the Lingones",
   "leucographis": "a plant used as a remedy for spitting blood",
   "leucon": "one of Actaeon's hounds",
   "leucothoe": "a daughter of Orchamus, king of Babylon, and Eurynome",
+  "leucus": "the Leuci, a people of Gallia Belgica near the Lingones",
   "levatus": "raised, lifted up",
   "levis": "light, not heavy; trifling, fickle",
   "lex": "a law, statute",
@@ -1311,6 +1316,7 @@
   "otium": "leisure, idleness, rest",
   "otium2": "leisure, ease, rest, idleness",
   "paeniteo": "to repent, regret, be dissatisfied",
+  "pagus": "a district, canton, commune",
   "palam": "openly, publicly (adv.); in the presence of (prep.)",
   "palamedes": "a son of Nauplius, a Greek hero before Troy, falsely accused by Ulysses",
   "palatinus": "of or belonging to the Palatine hill",

wiktionary_pron/utils/core_gloss.json

@@ -1374,6 +1374,13 @@
     },
     "note": "core_gloss"
   },
+  {
+    "lemma": "itis",
+    "expect": {
+      "contains": "you (pl.) go"
+    },
+    "note": "Caesar gloss audit (M-023l)"
+  },
   {
     "lemma": "jam",
     "expect": {
@@ -2996,6 +3003,13 @@
     },
     "note": "core_gloss"
   },
+  {
+    "lemma": "barba",
+    "expect": {
+      "contains": "beard"
+    },
+    "note": "Caesar gloss audit (M-023l)"
+  },
   {
     "lemma": "caesar",
     "expect": {
@@ -4758,6 +4772,13 @@
     },
     "note": "core_gloss"
   },
+  {
+    "lemma": "leucus",
+    "expect": {
+      "contains": "the Leuci, a people of Gallia Belgica"
+    },
+    "note": "Caesar gloss audit (M-023l)"
+  },
   {
     "lemma": "ligdus",
     "expect": {
@@ -5437,6 +5458,13 @@
     },
     "note": "core_gloss"
   },
+  {
+    "lemma": "leuci",
+    "expect": {
+      "contains": "the Leuci, a people of Gallia Belgica"
+    },
+    "note": "Caesar gloss audit (M-023l)"
+  },
   {
     "lemma": "leucon",
     "expect": {
@@ -11208,6 +11236,13 @@
     },
     "note": "core_gloss"
   },
+  {
+    "lemma": "impetus",
+    "expect": {
+      "contains": "an attack, assault, onset, charge"
+    },
+    "note": "Caesar gloss audit (M-023l)"
+  },
   {
     "lemma": "distribuo",
     "expect": {
@@ -11971,6 +12006,13 @@
     },
     "note": "core_gloss"
   },
+  {
+    "lemma": "pagus",
+    "expect": {
+      "contains": "a district, canton, commune"
+    },
+    "note": "Caesar gloss audit (M-023l)"
+  },
   {
     "lemma": "misereo",
     "expect": {