fix(voice): leave a term the vocabulary spells two ways by case as heard

[mrgoonie]

Fixes #590

Problem

The session vocabulary kept one spelling per lowercase key. With both UserService (a class) and userService (an instance) in a session, the transcript normaliser re-cased a heard term into whichever spelling won the key, turning one real symbol into the other.

Change

• buildRecognitionContext (packages/voice-adapters/src/coding-vocabulary.ts) keys terms by their exact spelling, so every spelling the session itself uses is kept. This is a behaviour change only when the session spells a word two or more ways by case.
• Glossary against session, per lowercase key:
  • the session spells the word one way only (a repository clarkcant, a provider gemini, a tool eslint): unchanged from main. The glossary entry and the session term contest by weight; the heavier spelling stays, a tie keeps the session's. So ClarkCant, Gemini and ESLint survive when the session term is not first in its list, and a dependency react still keeps react.
  • the session spells the word two or more ways by case (Voice normalizer: do not re-case a term whose spelling is ambiguous by case #590): new. Both session spellings stay and the glossary adds no third spelling beside them.
• The normaliser (transcript-normalizer.ts) groups terms that differ only by case:
  • heard in that shared spelling (userservice, UserService, userService), the span keeps its heard case and is reported as a known technical term: no change, no abstention, so no pointless retry;
  • heard as spoken words (user service), it abstains with every spelling as a candidate instead of picking a case;
  • such a span heard in a session spelling still counts as technical evidence, as a single term did.
• Unambiguous terms keep the feat(voice): restore the canonical casing of an exactly matched vocabulary term #574 behaviour (RedactSecrets -> redactSecrets in the same vocabulary).
• ADR-003 casing rule updated in EN and VI as a pair.

Behaviour changes compared with main

Only where the vocabulary holds two or more spellings of a word that differ by case:

• such a term heard in its shared spelling is left as heard (it was re-cased into one of them);
• its spoken words abstain with all spellings as candidates (they were written in one of them);
• a near match to such a term also abstains. Example: models Playwright plus tool playwright, "sửa playwrigt test" gives Playwright on main and abstains with both spellings here. A one-letter slip is no stronger evidence of case than the exact words, so abstaining is consistent and safe.

Everything else, including the glossary-against-single-session-spelling contest and the vocabulary sent to the recognizer in that case, is unchanged.

Evidence

Regression tests in packages/voice-adapters/test/transcript-normalizer.spec.ts:

• the vocabulary keeps UserService and userService, each exact spelling once, and adds no glossary spelling beside a word the session spells two ways (PNPM + Pnpm);
• the weight contest: a tool PNPM outweighs the glossary's pnpm (and its aliases go with it); a repository clarkcant listed after another project leaves the glossary's ClarkCant with its aliases;
• "sửa clark cant trước" and "Clarkcant build lỗi" with repositories: ["other-app","clarkcant","web"] give ClarkCant;
• "gemini model" with providers: ["google","gemini"] gives Gemini model;
• "eslint lỗi" / "Eslint lỗi" with eslint after 30 other tools give ESLint lỗi;
• userservice, UserService, userService and USERSERVICE are left as heard, with no change and no abstention, in Vietnamese and English sentences;
• user service abstains with both candidates;
• RedactSecrets is still re-cased beside the ambiguous pair.

The previous test "dedupes case-insensitively and keeps the glossary's aliases" only checked the CODING_GLOSSARY constant; it now asserts the built context in both directions of the weight contest. The new weight-contest tests fail against the previous head (71e9a44) and pass here.

Corpus bench (corepack pnpm --filter @clarkcant/voice-adapters bench:transcription, vi-en-coding-corpus.json, 84 utterances / 105 terms), output identical on main sources and this branch:

Stage	WER	CER	Technical Term Error Rate	Exact utterances	Changes	Abstained	Regressions
raw	22.8%	3.8%	77.1%	26.2%	-	-	-
normalized	3.7%	0.8%	16.2%	77.4%	64	1	0

Per kind (normalized) also unchanged: symbol 18/21, command 11/14, path 8/9, package 10/12, model 3/3, provider 6/7, acronym 9/9, glossary 19/24, version 2/2, branch 2/2, issue 0/2. Canonical references changed by the normaliser: 0. The corpus has no case-variant pairs and no multi-repository sources, so it exercises neither path; the unit tests above do.

The audio run (--audio ... --recognizer gemini-transcribe-live|gemini-live) was not run: external gate, no GEMINI_API_KEY in this environment.

Checks on head 4c00b7c: focused vitest packages/voice-adapters + apps/runtime/test/voice* (14 files, 231 passed), pnpm typecheck, pnpm invariants (14/14), eslint on the changed TypeScript files. pnpm verify was run on the previous head (71e9a44: 555 files passed, 1 skipped; 7363 tests passed, 37 skipped), not re-run locally on this head; CI covers it.

Notes

• Kept minimal around PR fix(voice): bias the dedicated recognizer only with terms that have no real neighbour #591 (fix(voice): the dedicated recognizer's vocabulary rewrites words the person actually said #573), which also touches transcript-normalizer.ts and ADR-003 (different hunks). With the weight contest restored, a session provider family such as gemini behaves as on main (it keeps Gemini when it is not first in its list; a lone gemini ties and stays lowercase, as on main), so fix(voice): bias the dedicated recognizer only with terms that have no real neighbour #591's case restoration still holds.
• Known limit, not changed here, tracked in fix(voice): prefer a run-together compact over a shorter two-word match ('clark cant web') #608: the matcher prefers a two-word form over a three-word run-together compact. When a session holds both ClarkCant and clarkcant, "clark cant web" abstains on "clark cant" (on main it became the also-wrong ClarkCant web) instead of becoming clarkcant-web.
• Official docs (clarkcant-web) do not describe the normaliser; no docs change needed.

[mrgoonie]

Review attestation: ready to merge at 4c00b7c8761e3c48b751694bbe0481488f312f21, reviewed by agent:code-reviewer.

A push to this PR makes this attestation stale; the new head needs its own review.