Restore saved-audio speaker recovery

[r3dbars]

Summary

• allow retained-audio re-transcription while speaker review is pending
• keep active dictation, recording, model-preparation, and background-work guards
• preserve focused coverage for recovery from a failed speaker-finalization state

Fixes #1681

Verification

• bash scripts/dev/agent-preflight.sh — passed
• bash run-tests.sh --filter RecentCaptureScannersTests — 133 passed
• bash run-tests.sh — 12,194 passed
• bash build-deps.sh --force — passed; existing compiler warnings only
• bash build.sh --no-open — passed; signature and launch smoke passed
• bash run-integration-smoke.sh — passed; core, wake, and recovery-merge smoke passed
• swift test — skipped; no Core/public SwiftPM seam changed

Independent review

• codex review --uncommitted was started against the true origin/main HEAD and stopped after approximately three minutes without producing a final verdict. Review timeout is recorded here rather than blocking the PR.
• Rejected findings: none; no completed reviewer findings were available to reject.

Manual real-device reproduction with the affected retained audio was not run and remains unknown.

[r3dbars]

Superseded by #1696. The replacement adds exact request-owner fencing, stale-finalizer isolation, retained-audio preservation through validated startup, and the required deterministic regressions. #1696 remains draft because the full fast-test gate and live affected-artifact proof are still open.