docs: M-031 live-verified, prayer harness into _sessions/, M-034/M-035, Pullfrog lesson

wiktionary_pron/NEXT.md

@@ -1,30 +1,33 @@
 # Next
 
-_Updated 2026-10-11 — branch `fix/macronizer-morpheus-wordform` (site PR #11, engine PR #2)_
+_Updated 2026-10-11 — fix merged (engine PR #2 `831cf70`, site PR #11 `81595ad`) and deployed_
 
 ## State
-Morpheus words no longer display as their lemma (`effuget → effugo`, M-031):
-engine `45e7821` + `5fbd757`, site `793d5c8` + sync. Full-prayer token test on
-the local page: wrong-word tokens 12 → 0, strict 99.14% → 99.49% (M-033).
-Green: engine jest 70/70, gold 98.86%, parity EXACT; site 2075 + 348, e2e 5/5.
+Morpheus words no longer display as their lemma (`effuget → effugo`, M-031),
+live-verified. Full-prayer token test on the live page: wrong-word tokens
+12 → 0, strict 99.14% → 99.49%, 11 of 16 remaining misses shared with
+Accenteur (M-033). Green: engine jest 70/70, gold 98.86%, parity EXACT;
+site 2075 + 348, e2e 5/5.
 
 ## Open threads
-- After merge + deploy: re-run `node F:/temp/lea/prayer/prayer-v4.mjs` (from
-  `wiktionary_pron/`, `PRAYER_URL` unset = live) and confirm the live numbers
-  match the local ones before quoting anything.
 - Latin SE 27176, answer 27340: moderator cmw commented "concerned about the
   2-3% difference". Reply with per-token strict numbers from M-033 (n=3152,
   revision-labeled), never mixed with the 1140-word figures. User posts it.
 - M-032: the page writes æ/œ back as ae/oe (176/192 ligature tokens) — mirror
   them from the input in `applyOrtho`, by position, like j/v.
+- M-034: wire `filterAccents` into `parseAnalysisLine` (Python parity), re-measure.
 - ISSUES H-002, H-003, M-023m `meneni` still need owner decisions.
 
 ## Running / unfinished
 - `python -m http.server 8000` (PID 12644 on 2026-10-11). Before killing, confirm
   the PID still owns port 8000:
   `Get-NetTCPConnection -LocalPort 8000 -State Listen` → `OwningProcess`.
 - Deploys are manual: `gh workflow run jekyll-gh-pages.yml --ref main`.
-- Prayer harness + Accenteur copies are scratch in `F:/temp/lea/` (FINDINGS.md).
+- Prayer harness + findings: `_sessions/2026-10-11-prayer-test/`; it reads the
+  Accenteur copies from `F:/temp/lea/accenteur/` (re-fetch steps in FINDINGS.md).
+- Not yet seen in a real browser: the stale-row clear on a RETURNING visit
+  (only unit-tested). Check in your own Chrome: console line
+  `stored Morpheus analyses are stale, clearing them`, then `effuget` → `effuget`.
 
 ## Don't redo
 - **Bare-lemma Morpheus lines echo the typed word** (Python postags.py 426-436);

wiktionary_pron/_sessions/2026-10-11-prayer-test/FINDINGS.md

@@ -0,0 +1,164 @@
+# Full-prayer token test — macronizer page vs Accenteur (2026-10-11)
+
+**Question.** The SE answer (latin.SE 27176, answer 27340) quoted "97-98% agreement"
+from a distinct-word list filtered to our wordlist; a moderator questioned the
+gap. What do the tools score on whole prayers, as a user runs them?
+
+**Files.** `prayer-v4.mjs` is here; the records JSONs and Accenteur copies are
+not committed. To re-run, re-fetch Accenteur's `static/js/data.js` +
+`accenteur_core.js` at `91823fe07b269189e8c0dd87a39caefdc1052c7d` (saved as
+`data_pinned.js`/`core_pinned.js`) and at master (`data.js`/`accenteur_core.js`)
+into `F:/temp/lea/accenteur/`, the path the script reads (edit it if moved); the
+Preface comes from the gregorio corpus (`corpus/missal/prefaces/`).
+
+**Method.** `prayer-v4.mjs` (run from `F:/projects/wiktionary_pron/wiktionary_pron`
+so Playwright resolves; `PRAYER_URL` = live or local page, `PRAYER_OUT` = records file).
+- Texts: the 13 Ordo Missae gold files (`latin-macronizer-wasm/test/data/accent-corpus`)
+  + the Preface "De Christo universorum Rege" (gregorio corpus). Input = the gold
+  text with acutes removed. Lines with no acute are rubrics and are skipped
+  (same rule as `test:accent`); 3152 gold-line tokens.
+- Ours: the real page in Chrome, "Mark long vowels" off, "Stress accents" on;
+  spans read per result row, aligned per line (merge-aware).
+- Accenteur: `accentify()` run in a Node VM at `91823fe` (README link,
+  `data_pinned.js`/`core_pinned.js`) and master (`data.js`/`accenteur_core.js`,
+  byte-identical to the live demo). Files in `F:/temp/lea/accenteur/`.
+- Scoring per token: accent position must equal gold; bare gold needs bare
+  output; unknown = miss. **Strict** also requires the same word (æ≡ae, j≡i).
+
+**Results** (`records.json` = live before fix, `records-local-fix.json` = local after M-031):
+
+| | before (live) | after fix (local) |
+|---|---|---|
+| ours, strict | 3125/3152 = 99.14% | 3136/3152 = 99.49% |
+| ours, raw (position only) | 99.37% | 99.49% |
+| wrong-word tokens | 12 | 0 |
+| Accenteur 91823fe, first / any | 98.79% / 99.43% | same |
+| Accenteur master, first / any | 98.22% / 99.33% | same |
+| same output as Accenteur 91823fe | 98.83% | 98.95% |
+| both-cover subset (3111): ours / Acc 91823fe | 99.39% / 99.13% | 99.52% / 99.13% |
+
+**Hand-verified:** the 12 wrong-word tokens (Morpheus bare-lemma lines, raw
+cruncher output captured from the page); the fix (unit tests on those lines,
+e2e on `effuget Adae`); the stale-render and alignment harness bugs (see LESSONS
+2026-10-11). **Not verified:** whether each remaining shared miss is a corpus
+error vs a quantity error (classified only for the 1140-word set, M-024/M-027).
+
+**Caveats.** Per-token and distinct-word figures have different denominators —
+never quote them together. The page expands æ/œ to ae/oe (M-032); strict scoring
+treats them as the same word.
+
+**Every miss, live after the fix** (3152 tokens; corpus-wide counts from all 871 files):
+
+Ours — 16 tokens:
+- Books contradict their own majority (our form = the majority) — 6: `Filii`
+  (vs `Fílii` 116×), `Omnipotens` (vs 101×), `igitur` (vs 14×), `pariter` (vs 22×),
+  `perséveret` (vs `persevéret` 2×), `indignátionem` (vs `indignatiónem` 6×).
+- Convention — 3: `Angelo`, `Alitur`, `Angelis`. The corpus never writes an acute
+  on a capitalized initial (Solesmes Option 2; 2 cases in 871 files); we do, by design.
+- Real errors — 7 tokens / 6 words: `effúget→éffuget` (Morpheus gives no vowel
+  length), `Parácliti→Paraclíti`, `líberet→libéret`, `agámus→ágamus` ×2,
+  `pervénit→pérvenit` (wordlist reading choice), `seípsum→seipsum` (split
+  into se+ipsum, Python-identical).
+
+Accenteur 91823fe — 38 tokens: the same 6 book contradictions; 11 unknown
+(`Kýrie` ×6, `Concedátque`, `vobismetípsis`, `astántes`, `misercórdiam` — a corpus
+typo for misericórdiam, `seípsum`); 21 wrong placements (`totíus→tótius` ×5,
+`ubíque→úbique` ×3, `véteris→vetéris` ×2, `mæstis→máestis` ×2 (two syllables, no
+accent), `María`/`Maríæ→Mária`/`Máriæ`, `eádem→éadem`, `cónditus→condítus`,
+`látere→latére`, `líberet→libéret`, `Parácliti→Paraclíti`, `effúget→éffuget`,
+`pervénit→pérvenit`). It follows Option 2, so the 3 capital cases are right for it.
+
+Accenteur master (live demo) — 56: gets `Kýrie`, `astántes`, `mæstis`, `pervénit`
+right, but loses `perveníre` ×4, `transíre` ×2, `servíre`, `Eiúsque`,
+`cælestiúmque` (unknown) and moves the accent onto the penult of `-ere`
+infinitives and similar (`redímere→redimére`, `ágere→agére`, `géneri→genéri`, …).
+
+## Full corpus (2026-10-11): 73,961 tokens, 345 files
+
+`corpus-run.mjs` (live page + Accenteur, resumable) → `corpus-analyze.py` →
+`corpus-analysis.txt`. Prose with independent accents only: Ordo, Prefaces,
+Solesmes psalms (`psalms/special`), antiphons, responsories, Regula (one copy),
+Rogationes. Excluded: hymns (metrical), `psalms/vatican` (accents generated
+with Collatinus = Accenteur's own data, circular), Breviarium O.P. (all hymns),
+duplicate Regula/Gloria. 0 lines unaligned.
+
+| | correct | capital* | books-inconsistent | unknown | real | unknown+real |
+|---|---|---|---|---|---|---|
+| ours, page | 99.12% | 264 | 33 | 15 | 335 | 0.47% |
+| ours, engine output | 99.14% | 264 | 33 | 0 | 339 | 0.46% |
+| Accenteur 91823fe | 99.19% | 0 | 33 | 127 | 441 | 0.77% |
+| Accenteur master (live) | 98.52% | 0 | 31 | 255 | 807 | 1.44% |
+
+73,755 tokens after excluding 206 non-words (2026-10-12, the user spotted
+`XXXVIII → Xxxviíi`): Roman numerals 72 (Regula chapter numbers etc.), the
+versicle/response markers `V.` 107 / `R.` 20, the name placeholder `N.` 7.
+`O`, `A`, `E` stay (interjection, prepositions). Before the exclusion:
+99.10 / 99.12 / 99.17 / 98.50%. The review's confidence intervals were computed
+on the unfiltered set; the shift is ~0.02 pp for every tool.
+
+"books-inconsistent" is per tool: the corpus has 41 places where the book
+writes a minority form of the same word (compared within the same initial
+case); a tool is counted there only if it wrote the majority form. At those 41:
+ours 33 majority / 2 matched the book (`décore`, `ínvenit`) / 6 other
+(`Mariam→Máriam`); Accenteur 91823fe 33 / 1 / 7. Two earlier classifier
+versions were wrong: lowercasing all votes let the capital convention ('Omnia'
+bare) outvote 'ómnia'; lowercase-only voting let 'mária' (seas) outvote the
+name 'María'.
+
+\* capitalized word the corpus leaves bare, accented on its first letter — the
+corpus applies Solesmes Option 2 with 2 exceptions in 871 files; we deliberately don't.
+With Option 2 ours would score 99.48-99.50% (Accenteur unchanged, it follows it).
+
+Per section ours leads in Ordo, Prefaces, antiphons; Accenteur leads in the
+Solesmes psalms (99.50 vs 99.15), Regula, responsories — the psalm lead is the
+capital convention (124 tokens): on substance, psalms ours-wrong/Acc-right 75,
+Acc-wrong/ours-right 94.
+
+Our largest fixable error groups (tokens): capital convention 265 (a decision);
+`Ierúsalem→Ierusálem` 56 (exception); compounds the tokenizer splits and so
+leaves unaccented — `Verúmtamen`, `semetípsum`, `idípsum`, `seípsum`,
+`unusquísque`, `semetípsis` ~36 (engine: stress the rejoined word); `Maríam`
+15 (extend the María exception); `Iesse→Iésse` 7 (2 syllables); `Basan→Basa`,
+`Hermon→Hermo` 9 (page display drops letters the engine keeps — M-032 family);
+`-ere` imperatives read as other forms (`Convértere→Convertére`, …), proper
+names (`Míchael`, `Ráphael`, `Andréas`), `érue` (missing from the wordlist).
+Scorer note: an œ+acute bug in an earlier analyzer version cost us 18 tokens;
+fixed (combining marks attach to the preceding letter).
+
+## Independent review of the full-corpus numbers (2026-10-11)
+
+A cold reviewer (separate agent, own code, `F:/temp/lea/review/`) recomputed
+everything from the raw records: all four accuracy figures, the class counts,
+agreement 98.63%, Accenteur any-candidate 99.72%, per-section numbers and the
+Option-2 counterfactual reproduce exactly (one token differs in class:
+`áëris`, a gold diaeresis). Gold token sequence matches the corpus text for
+all 73,961 tokens; no alignment shift. Exclusions verified. Corrections to how
+the numbers were described above:
+
+- **Raw is a statistical tie, not a lead either way.** Engine - Accenteur
+  91823fe = -0.05 pp; 95% bootstrap CI by file [-0.18, +0.13], by word type
+  [-0.44, +0.40].
+- **About 9% of tokens are repeated lines** (antiphon books overlap, Exsultet
+  brevior/longior). De-duplicated: Accenteur 91823fe 99.26% vs ours 99.10%
+  (99.43% with Option 2) — a small Accenteur lead raw.
+- **"0.48% vs 0.79% real+unknown errors" is Option-2-adjusted for ours only.**
+  Counting the capital-initial tokens as errors: ours 0.84% vs Accenteur 0.79%.
+- **The Option-2 lead (+0.32 pp) is likely real in direction but not
+  established**: CI by file [+0.18, +0.54], by word type [-0.03, +0.75]. It is
+  concentrated in a few frequent words (Accenteur: `ubíque` 88, `María` 80,
+  `Maríæ` 37 of its 583 misses; ours: `Ierúsalem` 56).
+- **"Option 2 without exception" was wrong**: 2 counterexamples in 871 files,
+  one scored (`Ángelum`, antiphons/Adventus.txt) — Option 2 is worth +264, not +265.
+- **Where ours is clearly ahead**: on the 1,771 tokens where Accenteur has 2+
+  candidates, its first choice is right 76.5%, ours 94%. Ours also guesses
+  unknown words (619 flagged, 76% right) where Accenteur abstains (142).
+  On tokens Accenteur knows: ours engine 99.17% raw (99.53% Option 2),
+  Accenteur 99.36%.
+- **Development overlap**: the Ordo files are the engine's gold test
+  (`test:accent`), and ours leads exactly in Ordo and Prefaces — in-sample.
+- Required wording: "accent-position accuracy on a Solesmes/Vatican prose
+  subset, with ae/oe respelling ignored".
+
+**Live after deploy** (`records-live-fix.json`, 2026-10-11, merges `831cf70`/`81595ad`):
+identical to the local run — strict 3136/3152 = 99.49%, 0 wrong-word tokens,
+16 misses, 11 of them shared with Accenteur 91823fe.