Overlay: a slow transcription holds the disc instead of bouncing

[dinooo13]

Summary

When a toggle dictation was stopped, the pill sometimes dived, came back up saying "Transcribing…", and dived again.

The cause. At the release the pill gathers into the disc and dives. If the text had not been pasted 300 ms after the release, a spinner task flew the pill back up as the "Transcribing…" row. At Quick the dive takes 460 ms (260 ms gathering, 200 ms slide), so every transcription between 300 ms and the end of the dive bounced. A release inside an engine pass is exactly that case: a 42.6 s toggle dictation logged release-to-paste 0.489 s … engine 0.434 … engine-time 0.179 s. Long toggle dictations run more passes, so they hit it more often.

Now there is one exit, decided where the gathering ends:

• Pasted by then (the usual case): the disc dives, as before.
• Not yet: the disc stays where it is, a spinner replaces the wave, and it dives when .idle arrives. The clipboard hint and an error open out of it in place.
• The late "Transcribing…" pill, its task and its 300 ms delay are removed. Only the polish path, which waits seconds, still shows the row.
• A flight-in that completes after the release has begun the exit no longer opens the disc back into the row.

OverlayDemo gains a "late" take, pasted 450 ms after the release, the reported case.

Test plan

• swift build, swift test: all suites pass.
• scripts/overlay-demo.sh, Compact at Quick. "late" (450 ms) and "slow" (900 ms) gather into the disc, hold it with the spinner and dive once; no "Transcribing…" row and no second flight. "plain" and "copied" look the same as before.
• Live: a long toggle dictation whose release lands inside an engine pass.

No benchmark: only the overlay changes, and nothing between recordingStopped and inserted in DictationCoordinator is touched.

🤖 Generated with Claude Code