docs(learnings): a phoneme dialect mismatch is invisible to every guard built to catch a bad G2P

The Russian G2P passed the inventory check (zero dropped codepoints, with a
control that fires on that exact path) and still lost the roundtrip to espeak,
because the two spell the same sounds differently and BOTH spellings are inside
zonos's inventory. The guard built to catch a bad G2P cannot see the one thing
that was actually wrong. Generalisation: when a model was conditioned on some
tool's output, correctness of your transcription cannot answer whether it is
the same spelling.

Also records the two instrument flaws the run exposed — a control anchored to a
baseline that itself collapsed to 0.000, and an espeak baseline below the
pre-registered floor making the roundtrip inconclusive rather than a pass — and
the disjoint-files discovery in the upstream data.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: CrispStrobe

LEARNINGS.md

@@ -11,7 +11,7 @@ If a lesson is still "live" (affects current work), it's linked from
 ---
 
 <!-- BEGIN INDEX POINTER — generated by tools/gen-learnings-index.py -->
-**Do not read this file top to bottom** — it is ~1 MB across 301 sections.
+**Do not read this file top to bottom** — it is ~1 MB across 302 sections.
 The generated topic + per-model index is [`docs/LEARNINGS-INDEX.md`](docs/LEARNINGS-INDEX.md).
 Find the lesson there, then jump with `grep -n "<heading text>" LEARNINGS.md`.
 Regenerate both with `python tools/gen-learnings-index.py` after adding a section.
@@ -18889,3 +18889,78 @@ before submission, and export the build cache immediately after a successful
 build as well as after final validation. The corrected P100 run was
 transcript-exact and made v0.23 0.53% faster on the 134 s clip; its 7,927-file
 cache became the next account-matched dataset version.
+
+## A phoneme dialect mismatch is invisible to every guard built to catch a bad G2P (#435 follow-on, 2026-09-15)
+
+Russian was the last language for which zonos still needed espeak-ng (GPL-3.0).
+A built-in G2P now covers it: an 812,953-entry dictionary with lexical stress
+already resolved, plus letter-to-sound rules. The inventory question that #435
+exists to police was answered up front and answered well — the dictionary uses
+37 distinct codepoints, 36 of them already in zonos's `conditioning.py` symbol
+list, and the exception was the upstream stress marker, a BACKTICK (U+0060) that
+is 7.57% of all symbols in the file. Rewriting it to U+02C8 when publishing the
+dictionary took the fit to 100%. Had that been missed, the built-in path would
+have produced audio, returned 0, and silently lost every stress mark.
+
+The Kaggle run confirmed it: **zero dropped codepoints on the built-in path**,
+and — the part that makes that number mean something — a control that fires on
+that exact path. A one-line dictionary carrying the original backtick was
+injected through `CRISPASR_RU_DICT_PATH`, and the counter reported
+`dropped=1 hist={U+0060: 1}`. Without that arm, "the built-in dropped nothing"
+is a negative the instrument has never been shown capable of contradicting.
+
+And the ASR roundtrip still preferred espeak, 0.585 to 0.293.
+
+Running espeak's own `ru` voice over 2,200 dictionary words locally said why,
+and it is not a defect in the G2P. The two transcribe the same sounds in
+different symbols:
+
+    ours    məɫɐkˈo i xlʲep lʲɪʐˈat na stɐlʲˈe v bɐlʲʂˈoj kˈomnətʲe
+    espeak  mʌɭʌkˈo ɪ xɭʲˈep ɭʲiʒˈɑt nə stʌɭʲˈe v bʌɭʃˈoj kˈomnʌtʲi
+
+Raw symbol agreement 57.7%; not one word of the 2,200 identical. **Every symbol
+on both sides is inside zonos's inventory**, so the drop counter — the guard
+built specifically to catch a G2P that hands a model something it cannot use —
+is blind to this by construction. So is the inventory unit test. So is the
+refusal path. A narrower, phonetically better transcription can be strictly
+worse for a model trained on a broader one, and none of the checks notice.
+
+This is #316 (Kokoro trained on misaki's spelling, 58% agreement from CMUdict)
+one language further on, which is the useful generalisation: **whenever a model
+was conditioned on some tool's output, the question is not "is my transcription
+correct" but "is it the same SPELLING", and correctness cannot answer it.** The
+fix has the same shape too — the G2P keeps emitting its own transcription and
+the consumer converts, because one dictionary serves consumers with different
+conventions.
+
+The conversion was DERIVED rather than assumed: each rule is the majority
+alignment over those 2,200 pairs. It takes agreement to 88.0% and exact word
+matches from 0.0% to 32.6%. One rule that felt obvious was measured and
+REJECTED — moving `ˈ` from before the stressed vowel to the syllable onset
+drops agreement to 75.1%, because espeak's ru voice marks the vowel too.
+Another, `tɕ`→`tʃʲ`, measured better (89.6%) and was deliberately held out of
+the shipped function so that the code matches the arm that was measured.
+
+Two further things the run taught about instruments rather than about Russian.
+First, a control leg compared the wrong-language arm's word-F1 against the
+baseline for sentence 0 — whose espeak score was itself 0.000, so the leg could
+not fire whatever the control produced. Its agreement leg fired decisively
+(0.063 against 0.63–0.77), which is the only reason the flaw was visible.
+**A control anchored to a baseline that can collapse is not a control.**
+Second, the espeak baseline for Russian was 0.585, below the pre-registered 0.60
+floor, and 0.000 on one of three sentences — zonos-v0.1 does not list Russian
+among its languages. So the roundtrip measures zonos's Russian more than it
+measures the G2P, and the honest verdict is INCONCLUSIVE for the default flip,
+not a pass and not a failure. Spanish was once passed for exactly the mirror of
+that reason.
+
+The data itself carried one more lesson, unrelated to measurement. The upstream
+project shipped `heteronyms.txt` alongside the vocabulary, and the two files are
+DISJOINT: 0 of the 17,359 heteronyms appear among the 812,953 entries. They are
+not entries with a chosen reading — they are the words upstream removed because
+it could not choose. "On a heteronym, take the dictionary's reading" describes
+an option that does not exist, and the set is not exotic (`все`, `уже`, `потом`,
+`чем`, `небо`, `тест`). **Check the relationship between a source's files before
+designing around it**; the shape that was assumed (a dictionary with one reading
+per ambiguous word) and the shape that shipped (a dictionary with those words
+deleted) lead to completely different fallback designs.

docs/LEARNINGS-INDEX.md

@@ -17,7 +17,7 @@ per-backend porting detail (Per-model port notes + the family cross-reference).
 Reach for the topic groups when you are stuck on a *technique*, and the
 cross-reference when you already know which model you are touching.
 
-## Index by topic (301 lessons)
+## Index by topic (302 lessons)
 
 **Security & untrusted input** (2)
 
@@ -357,13 +357,14 @@ cross-reference when you already know which model you are touching.
 - L18205 — Four bugs that together caused empty/garbage transcripts; all patched in src/vibevoice.cpp +…
 - L18430 — "Verified byte-identical at 225 s" verified a clip, not a length — and a decoder that drops spans needs a repair pass, not a better cap
 
-**Uncategorised** (5)
+**Uncategorised** (6)
 
 - L214 — An arm that fails under every condition cannot discriminate a fix
 - L235 — Hardcoded token ids need a guard that DECODES them
 - L286 — Model-emitted sentinels are not transcript text
 - L18657 — An ISA gate that "safely refuses" the only CPU module ships a process that cannot run — and every safety layer downstream assumed the layer above had fired
 - L18770 — An amplifying decoder needs a bound against INPUT SIZE, not duration — and the regression corpus that would have caught it had never been committed
+- L18893 — A phoneme dialect mismatch is invisible to every guard built to catch a bad G2P
 
 ## Cross-reference by model / family
 
@@ -407,6 +408,6 @@ lookup, not a partition.
 - **irodori** (2) — L1941, L15202
 - **f5-tts** (1) — L14544
 - **kokoro** (4) — L906, L7782, L15839, L16512
-- **zonos** (1) — L12209
+- **zonos** (2) — L12209, L18893
 - **omniasr / omnivoice** (6) — L733, L1636, L5103, L14341, L14369, L14755
 - **glm / mimo / mega** (3) — L11557, L14252, L14309