[diegosouzapw]

Summary

• add conservative audio-capability resolution with registry, static, and synced-catalog precedence
• add the priority-6 Audio Bridge guardrail, safe audio extraction, authenticated local transcription self-loop, cache/stats metadata, and partial-failure preservation
• make the Modality Bridge Audio tab functional and source transcription models from the authenticated unified model catalog
• translate the 12 new UI keys across all 42 non-English locales and document the runtime contract

Integration state

Parent PR #9782 is merged. This PR is now based directly on release/v3.8.50 and was rebuilt as the same eight audio-only commits on release SHA 4fc04a56045e0afb9b81b38ab8dc74fa9983e89d.

• PR head: cdea9afd4223b8fc54c12e1601b8fe239706fbf8
• diff: 67 paths, matching the pre-retarget audio diff
• all 12 Audio Bridge values were structurally reconciled and preserved in all 43 locale catalogs
• branch and remote head are clean and identical

Verification

Passing on the final retargeted head:

• focused Audio Bridge, settings, capability, catalog-route, and glossary Node suites: 49/49
• focused Modality Bridge UI suites: 3 files, 15/15
• 17 Vision Bridge canary files plus repro fix(providers): Vision Bridge describe-model unreachable when no openai provider connected → raw image forwarded to non-vision model #8430: 138/138
• npm run test:vitest: 37 files, 344/344
• npm run typecheck:core
• npm run lint
• npm run check:cycles: 406 files, no cycles
• all 42 locale validators
• npm run i18n:sync-ui:dry -- --translate-markers: 0 additions, 0 translations
• i18n coverage: 42/42 locales at 100% (11542/11542)
• value-drift and zh-CN/zh-TW glossary checks
• Prettier for all locale catalogs and git diff --check
• Build App run 31352954886: success on the exact PR head, including artifact upload
• Quality run 31353108429: change classification, merge integrity, Vitest, ESLint/ratchets, and unit shards 1/4 and 3/4 passed; shard 2/4 had only the inherited env/docs failure, while shard 4/4 had that same docs failure plus two stream assertions reproduced on the pure base; its duplicate advisory build remained queued when this description was refreshed

Base-red discrimination

The manual workflow_dispatch quality run exposed three inherited release reds. Each was reproduced on a disposable worktree at the pure base SHA 4fc04a56045e0afb9b81b38ab8dc74fa9983e89d; none of the affected paths is in this PR:

1. Docs sync: .env.example contains OMNIROUTE_STRICT_SYSTEM_PROVIDERS, while docs/reference/ENVIRONMENT.md does not.
2. File-size absolute fallback: the same four pre-existing violations in open-sse/handlers/videoGeneration.ts, open-sse/mcp-server/schemas/tools.ts, open-sse/mcp-server/server.ts, and open-sse/utils/stream.ts.
3. Open-SSE typecheck ratchet: the same two TS2339 errors in open-sse/handlers/videoGeneration/falHandler.ts.
4. Unit shard assertions: check-env-doc-sync and issue-7793-env-doc-sync-repro fail on the same missing env documentation. The two additional failures in stream-numeric-ids.test.ts and stream-utils.test.ts were run in isolation on both the PR head and the pure release SHA; both produced exactly 57/59 with the same replaced numeric ID and missing total_tokens assertion.

The file-size job used its documented absolute fallback because this was a manual dispatch and therefore had no pull-request base SHA. These reds are tracked as release-base health, including #9985; they are not assigned to the Audio Bridge diff.

Security and review reconciliation

The Standards and Spec reviews were reconciled before publication. Two findings were fixed: response-header CRLF hardening for both modality bridges and transcription-model discovery through the authenticated unified catalog. No remaining diff-owned code blocker was found.

Live Deepgram gate — merge blocked

Discovery was non-destructive and did not print credentials:

• 177 provider connections inspected locally
• 0 Deepgram connections and 0 active Deepgram connections with a usable key
• no DEEPGRAM_* environment variable or repository Actions secret name

The required real external input_audio -> text-only chat proof cannot be executed in the current environment. Per Hard Rule #18, this PR must remain open and must not be merged until that live gate is recorded.

The single unblock is to add or expose one usable Deepgram connection locally. Then send a known input_audio clip to a text-only chat target and record the HTTP status, x-omniroute-modality-bridge header, and a non-sensitive response excerpt. No VPS deploy was performed.

Related

• UI foundation: feat(dashboard): Modality Bridge settings page (vision tabs, model selector, stats, test button) #9782 (merged)
• video backlog: feat(backend): Video bridge (frame sampling + captioning) - backlog da série Modality Bridge #9760

[diegosouzapw]

Final stacked-head validation for 689e1e0

• parent feat(dashboard): Modality Bridge settings page (vision tabs, model selector, stats, test button) #9782 refreshed to 60c0f19
• Fast Production Build passed in 17m22s, including Next.js app and CLI bundle plus artifact upload
• Semgrep Cloud passed
• post-refresh focused verification passed: 49 Node tests and 12 UI tests
• branch is open, non-draft, mergeable, and has no unresolved review threads

No code blocker remains in this stacked diff. The single external gate remains the explicitly documented live Deepgram call, which cannot be run until one usable local Deepgram connection is exposed. No VPS deployment and no merge were performed.

[diegosouzapw]

⚠️ This PR branch is currently held by an active worktree session (Rule #22b). Keep as deferred until ownership is released.

⚠️ Review held: this PR branch is owned by an active parallel worktree session. Keep it in defer until ownership is released and re-run this PR assessment.

[diegosouzapw]

⚠️ This PR branch is currently held by an active worktree session (Rule #22b). Keep as deferred until ownership is released.

⚠️ Review held: this PR branch is owned by an active parallel worktree session. Keep it in defer until ownership is released and re-run this PR assessment.

[diegosouzapw]

⚠️ This PR branch is currently held by an active worktree session (Rule #22b). Keep as deferred until ownership is released.

⚠️ Review held: this PR branch is owned by an active parallel worktree session. Keep it in defer until ownership is released and re-run this PR assessment.

[diegosouzapw]

Post-merge Deepgram live gate — passed on VPS .15

The required product gate was completed on the Local VPS against release f1eb0b835722ca744692834c84661f1bc8d564a6 plus the surgical runtime fix in #10229 (d4b8bd9cabe9ef35a6069bedf2a024acc996900a). No credential or sensitive request body is included here.

• real WAV generated on VPS .15
• chat target: openai/gpt-4.1-mini, verified text-only for audio
• STT model: deepgram/nova-3, active provider credential
• request path: real input_audio through /v1/chat/completions
• HTTP status: 200
• transparency header: x-omniroute-modality-bridge: audio->text;model=deepgram/nova-3;parts=1
• sanitized response excerpt: OmniRoute Audio Bridge live verification number 739.
• expected transcript match: true

The gate initially exposed a real runtime defect: Node native FormData was incompatible with the packaged Undici fetch installed by ProxyFetch, producing invalid multipart data in the self-loop. PR #10229 adds the TDD regression and explicit multipart Buffer serialization. The exact candidate was transactionally deployed and the gate above then passed.

Freeze evidence on VPS .15: 23/23 focused Node tests, 4/4 Audio Bridge UI Vitest tests, focused ESLint, typecheck:core, check:cycles, fragment validation and changelog anti-eat comparison all passed. See #10229 for the complete evidence and exact SHAs.