fix(api): record audio transcription/translation/speech requests in call_logs (#13544)

[diegosouzapw]

Closes #13544

Root cause

open-sse/handlers/audioTranscription.ts and src/app/api/v1/audio/transcriptions/route.ts
never called saveCallLog() (@/lib/usageDb), unlike every other proxied API surface
(embeddings, images, video, rerank, search). Successful and failed
/v1/audio/transcriptions requests were therefore invisible in Dashboard → Request Logs,
even though the upstream provider was dispatched correctly and the transcription was
returned to the client. The reporter's secondary point was also confirmed: the response
route hardcodes costUsd: 0 with a comment claiming "no text body available" — stale for
the OpenAI-compatible default handler, which already buffers the full upstream JSON body
(including a provider's usage:{type:"duration",seconds:N}) before returning it; it was
just never parsed for logging.

The identical gap exists in the sibling /v1/audio/translations and /v1/audio/speech
routes — same code shape, not user-reported, fixed alongside since it's the same class of
defect.

Fix

• Threaded policy.apiKeyInfo.{id,name} from enforceApiKeyPolicy() into
  transcribeWithModel() / translateWithModel() (and the postHandler closure for
  speech), matching how rerank.ts and embeddings.ts already do this.
• Added saveCallLog() on both the success and error branch of all three routes, with
  provider, model, connectionId, duration, apiKeyId, apiKeyName — mirroring the
  canonical shape in open-sse/handlers/embeddings.ts / src/app/api/v1/rerank/route.ts.
• Added peekDurationUsage() in the transcriptions route: a best-effort peek (via
  response.clone().json(), never touching the original stream) at a provider's
  usage:{type:"duration",seconds:N} body, persisted on the call_logs row's
  responseBody for audit ahead of a future per-second cost rule. No cost pricing engine
  is added here — costUsd stays 0 for transcription, matching the existing comment's
  actual (if previously mis-explained) behavior.
• Combo fan-out (handleComboChat → transcribeWithModel/translateWithModel per target)
  now receives the caller's apiKeyId/apiKeyName too, so combo-routed audio requests are
  attributed correctly as well.

Regression test

tests/unit/issue-13544-audio-transcription-call-log.test.ts — drives a real
POST /v1/audio/transcriptions through the actual route (provider-node resolution,
enforceApiKeyPolicy, upstream dispatch) against a mocked OpenAI-compatible provider.

RED (on unfixed release/v3.8.51 tip):

✖ #13544: a successful transcription through an OpenAI-compatible provider node creates a call_logs entry
  AssertionError [ERR_ASSERTION]: expected a call_logs row for the successful /v1/audio/transcriptions request
  (none was created — the transcription path bypasses the normal call-log pipeline, #13544)

GREEN (after the fix):

✔ #13544: a successful transcription through an OpenAI-compatible provider node creates a call_logs entry (32359ms)
✔ #13544: a failed upstream transcription request also creates a call_logs entry (410ms)
ℹ tests 2
ℹ pass 2
ℹ fail 0

The success-path assertions also confirm the row carries provider/model/status/
apiKeyId, and that getCallLogById() returns the persisted usage:{type:"duration", seconds:3} on responseBody (issue Validation Plan steps 4 and 5).

Gates run

• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> → 0 errors/warnings
• npm run typecheck:core → 0 errors
• node scripts/check/check-test-discovery.mjs → OK, 0 new orphans
• node scripts/check/check-cognitive-complexity.mjs → OK (1276 violations vs baseline 1437 — improved, not regressed)
• node scripts/check/check-file-size.mjs → no ✗ on any file this PR touches
• node scripts/check/check-complexity.mjs → did not finish within the session's available
  time due to heavy concurrent load from several other active sessions on the shared devbox
  running the same repo-wide scan in parallel (confirmed via ps); this PR's diff is a
  small, flat threading of a few extra parameters plus straight-line if/else saveCallLog
  calls copied from the already-passing rerank.ts/embeddings.ts pattern, so it is very
  unlikely to move this ratchet. Flagging for CI to confirm rather than asserting a result
  I couldn't observe locally.
• Full regression + sibling test files (tests/unit/issue-13544-audio-transcription-call-log.test.ts,
  audio-transcriptions-combo-resolution.test.ts, audio-transcription-handler.test.ts,
  audio-transcription-openrouter.test.ts, audio-transcription-opus-filename.test.ts,
  audio-translations-combo-resolution.test.ts, audio-translations-route.test.ts,
  audio-speech-dynamic-node-9096.test.ts, audio-speech-fishaudio.test.ts,
  audio-speech-handler.test.ts, audio-speech-ogg-alias-10587.test.ts) → 88/88 pass, no
  regressions.

Existing tests aligned

None — no existing assertion encoded the old buggy contract; all pre-existing audio
test files pass unmodified.

Not covered here

• No new per-second cost/pricing rule for transcription duration (out of scope per the
  triage plan — costUsd stays 0, unchanged behavior; only the upstream usage value is
  now persisted for future consumption).
• Non-default transcription provider formats (Deepgram, AssemblyAI, Gladia, etc.) get the
  core saveCallLog fix but not the duration-usage enrichment, since only the default
  OpenAI-compatible branch exposes a buffered JSON body to peek at.