Move a mic back to the engine after two empty pinned takes

[r3dbars]

Why

Before: in 1.1.66, #1857 moved built-in and wired mics onto the pinned Core Audio recorder, only to make dictation start faster. On one M1 MacBook Air (macOS 27), every held dictation on 1.1.66 came back empty: six were "no speech" and one was "signal but no words". The same Mac dictated fine on the engine path in 1.1.64, minutes before it updated. The person kept pressing the key, and those short taps showed up in Sentry as APPLE-MACOS-2J ("Dictation ended before the microphone finished opening"). PostHog shows the same Mac going 0 for 37. On the "no speech" takes the words are lost, because nothing is pasted and the audio isn't kept.

After: a mic that is on the pinned recorder only for speed, and gives two empty held takes in a row, goes back to the engine path it used in 1.1.65 for the rest of that app version. An update tries the recorder again. The switch sends one dictation_pinned_microphone_fell_back_to_engine event with stage empty_takes, plus the input's format and bucketed health counts, so we can see what the recorder saw on a Mac like this.

The recorder's own health checks found nothing wrong on that Mac: no restart, no silent-input flag, no fallback. So the cause inside the recorder is still unknown (n=1). This PR is the safety net, plus the data to find the cause.

Product Impact

• Affects: dictation
• Lane: dictation reliability
• Why this matters: on the one Mac we know about, 1.1.66 turned dictation into silent failures. With this change, a Mac like that loses at most two takes per version before it's back on the path that works.

What changed

• New Sources/Speech/PinnedDictationSpeedPath.swift scores each held take on a speed-only pinned mic as words or empty. recordingTooShort, modelFailure and cancelled takes aren't scored. It stores a per-mic, per-app-version count in UserDefaults, keyed by UID.
• PinnedDictationInputPolicy.recorderIsNeeded skips the speed path for a mic that was moved back. Bluetooth-default and user-chosen overrides still always use the recorder. New recorderIsSpeedOnly.
• DictationStartCuePolicy plays the click after recording starts for a moved mic, because the engine start is slower again.
• ParakeetPinnedMicrophone keeps the take's format and health counts. STTRouter.transcribe(preparedRecording:) scores the take after transcription. Starting a new take clears any stale marker.
• Analytics: new empty_takes stage. dictation_pinned_microphone_fell_back_to_engine also carries input_channels, input_rate_hz, pinned_mic_restart_bucket, pinned_mic_gap_bucket and pinned_mic_dropped_callback_bucket, and analytics-events.psv is updated. No device names or UIDs leave the Mac.
• New tests in PinnedDictationSpeedPathTests, plus new suites in the input-policy, start-cue and analytics-forwarding tests.
• Also updated: the run-tests.sh and scripts/dev/test-parakeet-lifecycle.sh source lists, the Speech/UI CLAUDE.md files, and the pinned-mic bullet in the privacy doc.

How I checked it

• scripts/dev/agent-preflight.sh
• bash scripts/dev/linux-checks.sh (runs without Swift): 48 passed, 0 failed
• Selected checks from .agents/test-matrix.yml for the files changed (CI green on 1381c6e)
• bash build.sh --no-open (CI app-build)
• bash run-tests.sh (CI checks)
• Performance budget passed
• bash run-integration-smoke.sh (CI spm-tests; its Parakeet lifecycle smoke compiles the input policy)
• swift test: not needed
• bash run-e2e-smoke.sh / swift test --package-path Tools/<Package>: not mapped
• Manual check:

Checks I could not run, and why:

• This session has no Swift toolchain, so the build and tests ran in CI only. check-analytics-emitters.py, the taxonomy normalizer and check-telemetry-keys.py passed as part of linux-checks.

Mac or hardware test still needed? If yes: the steps, the log line or UI change each step should show, and whether each step fails on main:

• An M1 Air is best, but any Mac with a built-in mic works for the fallback path. Hold the dictation key twice while staying silent, then hold it a third time and speak.
  • The second silent take should log the empty_takes fallback warning.
  • The third take should use the engine path (the backend isn't pinned_ioproc) and paste the words, with the click after recording starts.
  • On main, the third take stays on the pinned recorder.
• Restarting with a different CFBundleShortVersionString should use the pinned recorder again.

Risk Review

• Privacy / local-first behavior reviewed (the store is local; the event carries only counts and format)
• New analytics properties or Sentry tags avoid the sanitizer's drop fragments
• Checked the text-pin tests for every file I edited (check-source-pins.py --changed-only: PASS)
• Storage path or migration impact reviewed (one new UserDefaults key, pinned-dictation-speed-path)
• Public-facing copy stays concrete and matches current product scope (no copy changes)
• Release/update impact reviewed (the move resets on every app version by design)
• Agent PRs got an independent deep review of the full diff (READY at 1381c6e; its one finding, in the privacy doc, is fixed)
• UI changes include sanitized .agent-review/visuals/ evidence: no UI changes
• No private transcripts, audio, tokens, personal paths, or customer data are included

Notes

• A user whose built-in mic works fine can also get moved, if they hold the key and say nothing twice in a row. They lose about 0.2 s of start speed for that version, and nothing else.
• Engine takes aren't scored, so a moved mic stays moved until the next version.

Agent handoff

COORD_DONE: GREEN | this PR | adaptive fallback from the pinned recorder after 2 empty speed-only takes + format/health telemetry | none | none | linux-checks, source pins, CI | ship in 1.1.67

🤖 Generated with Claude Code

https://claude.ai/code/session_016DYGa1i8HDv497ewpWgaCM