[smolpaws]

Summary

Stacked on the existing Voice draft #5180. Companion App PR: enyst/insider#4. Bead smolpaws-vstale.

• Fence pending spoken results after newer speech.
• Avoid speaking another saved turn's result when typed input arrives during submission or completion.
• Keep saved answers intact; do not infer playback completion from queued speech.

No API schema, Codex pin, tool permissions, approval handling or run-admission changes. Steering is a separate follow-up.

Verification

52 relay/router tests, 1 existing live-socket voice route test, Ruff, targeted Pyright, and diff checks pass. New regressions failed before implementation. Protocol fixtures use a real subprocess, not a live provider. No microphone, login, paid calls or deployment.

Evidence and protocol limits. The pinned protocol cannot revoke queued speech or identify its input; browser local suppression remains necessary.

[smolpaws]

@coderabbitai review

[smolpaws]

Codex source re-review found a delivery race: newer input could arrive while appendSpeech waited for the subprocess write lock. Added a failing real-child-protocol regression, then moved the final generation check inside the lock immediately before the write. Superseded speech is not sent; saved work is unchanged. Now 53 relay/router tests pass; Ruff and targeted Pyright pass. No live provider or deployment. Transcript boundaries and the V3 browser interruption signal remain separate follow-ups.