[diegosouzapw]

Summary

Reconcile Video Bridge's failed/skipped-video transcript fallback with the strict
OpenAI content-part field allowlist. This preserves skygunner's compatibility
fix in #16395; it does not widen or revert that allowlist.

• Valid caller transcript/audioTranscript cues become supported text parts
  with explicit untrusted-client and unverified-media/timestamp warnings.
• Validate both tracks' combined input cue/UTF-8 budget before deduplication.
  Request-provided provenance never becomes server-authenticated provenance.
• Remove transcript extension fields from fallback video carriers and nested
  carrier objects; retain capability-safe video/marker behavior.
• Append structured redaction shadows without shifting prior shadow anchors;
  retain original-field, pending/log/artifact and cache/Memory protection.
• Keep native-supported and disabled bypasses unchanged; no new media/provider
  requests, settings defaults, promotion thresholds or operational-skill changes.
• Update the English contract to promise semantic text preservation, not arbitrary
  extension-field passthrough. Related to feat(backend): video-bridge #12150 P2 — remaining transcript-retention surfaces (raw snapshot, continuation, derived-prompt dispatches, response/semantic-cache) #12430 and feat(video-bridge): redact/retain-safe transcript data across logs, Memory and continuation #12150; not closing either.

Evidence

All tests executed on the designated VPS. HTTP RED reproduced lost top-level
transcripts and leaked nested extensions. GREEN initially passed the public
route/actual translator/SQLite retention matrix. Two-axis review then found a
duplicate-cue combined-budget bypass; new HTTP tests reproduced both count and
byte variants RED, then passed after correction. Native/disabled and Responses
controls also pass. Final focused HTTP result: 13 passing, zero failures/skips.

At exact clean head fd62a801a7eeb535a7268ffeb2b3a137bb64d638, 52 scoped
files / 482 tests PASS, zero failures/skips, including real FFmpeg and
strict translator tests. Docs-all, changed-file ESLint, core typecheck and
cycles ratchet (14 <= 14) PASS. Before/after SHA and cleanliness match.

File-size remains FAIL with four signatures reproduced byte-for-byte on exact
base e369ffc633131ef0b1ddbee9d2b40475e63c0808: responseSanitizer 1216 > 1214,
error.ts 1218 > 1217, translator-openai-responses-req.test.ts 1476 > 1470,
vscode-token-routes.test.ts 1268 > 1267. They are BASE-RED, not waived.
Strict translation drift FAIL on the new GUARDRAILS English source. It also
reports 396 stale locale records; warning-only reporting is not proof of fresh
translations. Overall scoped gate receipt is FAIL, not release-ready.
Static Spec and Standards rechecks report no remaining findings in the final
delta. Dependencies remain borrowed, not clean-lock certification.

Admission

⚠️ base-red inherited: #15306. Strict documentation translations are pending for
the new English contract; ongoing older-source translations cannot certify it.
Required full gates, original campaign-PR integration, provider/corpus evidence
and deployment remain separate. Draft; no baseline weakening or admin merge.

[diegosouzapw]

Continued the owner-approved video closeout on this original PR. Published reconciliation commit 1db721cd2d4ad157b84c62ea3d5ace708a09b28f (base eafe72ed415e6cdf9a98e628a38ef1f278830da6) by normal fast-forward; the original local worktree remains untouched.

Scoped evidence, using matching-lock borrowed dependencies (diagnostic, not clean-install certification):

• Initial HTTP matrix run hit its unchanged 120-second parent timeout during cold initialization; retain that FAIL/cancellation receipt.
• Serial repetition of the same candidate/command: 13/13 PASS, no skips/cancellations, 70.6 seconds.
• Seven affected Node regression files: 86/86 PASS, no skips/cancellations, 111.3 seconds.

Reviewed bounded fallback text conversion, client/unverified provenance, combined raw-input budgets, removal of transcript extension fields, caller-request immutability and retention shadows. No additional production code or timeout changes were made in this continuation.

Outstanding: real translations of the changed video contract, reconciliation against the newer release 94159e1b7536783aec456d4f1a1c02cf2b74403c, clean/full applicable gates and exact-candidate admission. This PR remains draft; focused PASS does not erase the first timeout or constitute merge approval. No queue label, merge, paid/provider-real call or deploy was performed.

[github-actions]

CI Coverage Report

• Coverage job: skipped
• PR test policy: success

Coverage artifact was not available for this run.