fix: resolve audio provider short-alias prefix in parseAudioModel (#10586)

[diegosouzapw]

Closes #10586

Root cause

GET /v1/models advertises both the short provider-alias id and the canonical
provider id for every model (default MODELS_CATALOG_PREFIX_MODE="dual"), e.g.
el/eleven_multilingual_v2 (alias el for elevenlabs) alongside
elevenlabs/eleven_multilingual_v2. parseAudioModel() in
open-sse/config/audioRegistry.ts only prefix-matched against the
AUDIO_SPEECH_PROVIDERS / AUDIO_TRANSCRIPTION_PROVIDERS / AUDIO_TRANSLATION_PROVIDERS
canonical registry keys, never against the short alias each provider is
registered under (src/shared/constants/providers/audio.ts) — so the
alias-prefixed id the catalog itself just advertised 400ed with "Invalid speech
model" when sent back to /v1/audio/speech or /v1/audio/transcriptions.

Fix

Added a Phase 1.5 prefix-match in parseAudioModel() that additionally tries
getProviderAlias(providerId) (from src/shared/constants/providers.ts —
covers AUDIO_ONLY_PROVIDERS, unlike the chat-only PROVIDER_ID_TO_ALIAS map)
for every registry entry, resolving to the same canonical providerId a
request already reaches via the long-form id. This only widens what parses; it
cannot change routing for any currently-working id. Verified there is no alias
string that collides with another provider's canonical id across the speech,
transcription, and translation registries.

Regression test

tests/unit/audio-alias-prefix-10586.test.ts — proven RED before the fix
(el/eleven_multilingual_v2 resolved to {provider: null, ...}), GREEN after.
Also adds enumeration tests covering every provider with a registered short
alias in AUDIO_SPEECH_PROVIDERS, AUDIO_TRANSCRIPTION_PROVIDERS, and
AUDIO_TRANSLATION_PROVIDERS (elevenlabs/el, deepgram/dg, soniox/sx,
assemblyai/aai, etc.) so both the alias-prefixed and canonical-prefixed
catalog ids resolve to the same provider.

Scope

Out of scope per the plan-file (tracked separately under the parent tracker
#10591): the missing /v1/realtime route, the
mistral/voxtral-mini-transcribe-2507 vs nested-openrouter-model mismatch, and
the google/gemini-2.5-flash-preview-tts vs vertex-only registry gap.

Gates run

• node --import tsx/esm --test tests/unit/audio-alias-prefix-10586.test.ts — 4/4 pass
• Full existing audio-module test suite (speech/transcription/translation handlers, provider-node selection, soniox, fishaudio, openrouter nested models, combo resolution) — 94/94 pass
• node scripts/check/check-file-size.mjs (changed files) — OK
• node scripts/check/check-changelog-integrity.mjs — OK
• npm run typecheck:core — OK
• npx eslint --suppressions-location config/quality/eslint-suppressions.json (changed files) — clean
• node scripts/check/check-complexity.mjs / check-cognitive-complexity.mjs — these are whole-repo ratchets (ignore file args); attempted twice but did not finish within the session due to severe devbox CPU saturation from many concurrent parallel sessions running the same full-repo scan simultaneously (confirmed via ps, not specific to this change). The diff is a ~10-line addition mirroring the pre-existing Phase-1 prefix-match loop in the same function, so complexity risk is minimal.

⚠️ base-red inherited: #9985