Summary

Keep the durable imported-audio journal until the active transcription reaches a terminal state.

User impact

The import journal was deleted as soon as model preparation handed work to the in-memory transcription task. A crash or force-quit during inference could therefore leave the scratch audio with no recovery journal or failed-queue record.

Fix

• Move journal cleanup behind a task-terminal callback.
• Invoke it only after transcript side effects commit, or after a normal failure/cancellation path has handled the scratch file.
• Keep the journal while inference is in flight so the next launch can recover the import.
• Add a blocking-engine regression test that asserts the callback waits for transcript commit.

Evidence and scope

• Confirmed against fresh origin/main at 2f8e01ef.
• The durable queue-journal work already on main covers model preparation; this closes the active-import handoff gap.
• Open PR Repair speech route recovery on current main #1569 covers speech-route recovery and does not overlap this change.

Verification

• swift test: 801 tests passed, 13 expected skips, 0 failures.
• bash run-integration-smoke.sh: passed.
• bash build.sh --no-open: signed app build and launch smoke passed; launch time 1352 ms against a 3000 ms budget.
• bash run-tests.sh: 13,046 tests passed, 0 failures.
• Independent local adversarial review: accepted the finding as real, reproducible from the current-main control flow, and safely fixable with this narrow callback handoff.

Proof boundary

The deterministic test covers the in-flight crash window, but no manual crash/force-quit or real hardware/audio proof was performed in this run.