[silverstein]

Stopping dictation during a blocking partial transcription can leave completed microphone chunks unread, cutting off the ending. Nearby typing can also produce a sound label such as (keyboard clacking) inside an otherwise correct transcript.

Stop and yield-to-recording now share one finalization path. Capture closes before its queued chunks are drained into the recovery recording and final decoder, without running more partial transcriptions during drain. Cancellation continues to discard audio and text, including cancellation during the final decode.

Dictation cleanup removes a narrow list of recognized bracketed or parenthesized keyboard/music labels and preserves the surrounding speech. Ordinary parentheses, nested content, quoted labels and explicit raw mode remain intact. Terminal mode preserves code formatting and ambiguous single-word tags.

Three queue/silence regressions cover a 2.4-second speech backlog, trailing silence, and silence without speech. Five sound-label regressions cover the exact observed phrase, adjacent labels, quoted/ordinary content, code/raw modes and malformed nesting. CI now executes the Whisper-enabled dictation and streaming decoder suites on all three platforms.

This branch includes the dictation control repair in #1105; those changes remain separately reviewable there. The PR targets main so the full required CI matrix runs. Rebase onto main after #1105 lands before merging this remaining change.

Validation: Mac Rust 1.95 formatting and strict workspace Clippy pass. The current sound-label cleanup suite passes 30 tests. The prior queued-audio candidate passed the required serial core suite (1,925 tests, one ignored), 27 dictation/cancellation tests and seven streaming decoder tests with the exact new CI filters. Its release CLI builds with parakeet + metal and reports dictation_addressed_stop_v1=true. Workflow lint, version sync and strict site release constants pass. Final serial core validation for the sound-label update passes 1,930 tests (one ignored).

Native limits: signed installed Dev identity and hotkey are verified; actual private MCP addressed-stop requests ended two native sessions while preserving the app. One microphone trial truncated the phrase and a second retained the complete ending, so repeatable native accuracy remains unverified. The second result also pasted into a changed active app while the user typed. Microphone/desktop tests are paused. The new sound-label filter is source-tested and has not been installed or qualified in the native app. Actual ChatGPT capture remains pending; this is a draft.