Count a long pinned take with almost no audio as empty

[r3dbars]

Why

Before: #1880 moves a mic off the pinned recorder after two empty held takes in a row. Empty means "no speech" or "signal but no words". A recorder that delivers no audio at all ends a take as "recording too short" instead. The scoring skipped those as taps, so a recorder that never delivers audio would never trip the fallback.

After: a "recording too short" take counts as empty when the recorder was open for at least 1.5 s and less than 0.5 s of audio arrived. The 1.5 s is the same line isAccidentalStart already draws between a mis-tap and a stalled mic. Taps under 1.5 s still aren't scored. A healthy mic that started slowly still delivers most of its hold, so it isn't scored either.

This follows up the 1.1.66 empty-dictation reports. A PostHog check found 3 of 8 Macs with zero working dictations on 1.1.66:

• Only one of them, the M1 Air from Move a mic back to the engine after two empty pinned takes #1880, had held takes that failed, and Move a mic back to the engine after two empty pinned takes #1880 already covers its "no speech" takes.
• On the other two, every failure was a key press under 0.25 s (Sentry APPLE-MACOS-2J pending_bucket lt_100ms / 100_249ms), so the mic had nothing to do with it.

This PR closes the one remaining gap that could hide a dead recorder.

Product Impact

• Affects: dictation
• Lane: dictation reliability
• Why this matters: on a Mac where the recorder opens but delivers nothing, the person sees "too short" on every take and never gets moved back to the engine. With this change they're moved after two.

What changed

• PinnedDictationSpeedPath.outcome gains heldSeconds and audioSeconds. recordingTooShort is .empty when held is at least accidentalStartMaximumPress (1.5 s) and audio is under deadRecorderMaximumAudio (0.5 s). Otherwise it's nil.
• ParakeetPinnedDictationRecording.startedUptime is reset once the capture's start returns. Stop reads the time before the drain, so neither a queued start nor the drain counts as time the recorder was open (deep-review S1).
• PinnedDictationSpeedPathTake carries heldSeconds and audioSeconds. audioSeconds is the take's recoveredRecordingTimeline.totalDurationSeconds (deep-review S2).
• Tests and the Speech CLAUDE.md bullet are updated.

How I checked it

• bash scripts/dev/linux-checks.sh (runs without Swift): 48 passed, 0 failed
• python3 scripts/dev/check-source-pins.py --changed-only: PASS
• bash build.sh --no-open (CI app-build, green on c45538f)
• bash run-tests.sh (CI checks, green on c45538f)
• bash run-integration-smoke.sh (CI spm-tests, green on c45538f; the Parakeet lifecycle smoke compiles this file)

Checks I could not run, and why:

• This session has no Swift toolchain, so the build and tests ran in CI only.

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:

• There's no easy way to make a healthy mic deliver nothing, so the logic is covered by the unit test. On a normal Mac, holding the key for 2 s or more while speaking should still paste, and a tap should still show "Hold the key while you speak".

Risk Review

• Privacy / local-first behavior reviewed (no new telemetry)
• New analytics properties or Sentry tags avoid the sanitizer's drop fragments (none added)
• Checked the text-pin tests for every file I edited
• Storage path or migration impact reviewed (same UserDefaults key)
• Public-facing copy stays concrete and matches current product scope (no copy changes)
• Release/update impact reviewed
• Agent PRs got an independent deep review of the full diff (READY at c45538f)
• UI changes include sanitized .agent-review/visuals/ evidence: not applicable, no UI changes
• No private transcripts, audio, tokens, personal paths, or customer data are included

Notes

• Follow-up to Move a mic back to the engine after two empty pinned takes #1880.

Agent handoff

COORD_DONE: BRIEF | this PR | long pinned takes with almost no audio now count toward the fallback | none | merge for 1.1.67 on a person's go | linux-checks, source pins, CI green | deep review READY

🤖 Generated with Claude Code

https://claude.ai/code/session_016DYGa1i8HDv497ewpWgaCM