fix(video): protect retained raw transcript fields

[diegosouzapw]

Raw video transcript/audioTranscript fields could remain in persisted request details and disk artifacts when video description failed or maxVideos left a sibling video unchanged. Mark these requests as observed even with zero replacements; after matching rendered-description shadows, apply the existing structural redactor to raw siblings on the retained copy. Preserve the provider payload and ordinary non-video retention.

Defined empty strings and null intentionally activate the same conservative retention/cache/Memory policy; absent fields and non-video metadata do not. Document this bounded contract in GUARDRAILS and its 66 translations. The source entry was fully synchronized on the base, so only its 199 source/target/timestamp fields are updated; all other entries and text outside the inserted paragraphs are preserved.

Refs #12430 (item 5, unreplaced raw-video residual). This is a partial delivery. The tracker remains open for its broader end-to-end acceptance; related STT/provenance work in #15780/#16061 is unchanged.

Validation:

• Two separate TDD reproductions through the real Chat Completions HTTP route and temporary SQLite: failed description with unknown native video capability, and one successful shadow plus an over-limit raw sibling. Each RED→GREEN pair keeps its fixture hash. The final fixture passes 6/6 (five subcases plus parent), including empty/null fields and ordinary-retention controls. It asserts the live upstream video parts remain structurally identical while retained detail/artifacts omit sensitive content; pending state is a preserved control, not an additional reproduced leak.
• Related selection: 21 files, 276 tests, 263 pass and 13 fail exclusively in the legacy HTTP fixture blocked by its host-without-port MockAgent allowlist. Its exact file hash matches the independent base RED for fix(tests): allow the exact retention fixture host through MockAgent #16122, which owns that fixture correction. No assertions or fixture from that PR were changed here. The final six-case rerun covers the later characterization addition; the selection is not described as entirely green.
• Core/dashboard typechecks, canonical focused ESLint/Prettier, file-size, both complexity gates, test discovery, changelog integrity and docs-all pass. The separate i18n drift output is byte-for-byte equal to the base result (24 source, 132 target, 132 stale, zero missing); no new drift.
• Independent SPEC and Standards review. The last follow-up changes only a comment explaining the extra structural clone. The 66 translations have technical/parity validation, not native-speaker review in every language.

The HTTP adapter runs real route/guardrail/logging code with synthetic description/inference and isolated DB/plugins, blocking external sockets. It does not exercise ffmpeg/STT, a real provider account, a full Next server, or every external store/plugin sink. No merge or deployment is performed here.

[github-actions]

CI Coverage Report

• Coverage job: skipped
• PR test policy: success

Coverage artifact was not available for this run.