[kitlangton]

Why

Long dictations were slow on a busy Mac: the Parakeet decoder ran 8 spin-waiting CPU threads, so under build load a 30 s dictation could take 12.8 s to transcribe, and nothing was transcribed until release. Meetings existed only as a debug prototype, with echo duplicating remote speech as You on speakers and no way for agents to read past calls. This PR ships both as Hex 2.2.0.

What Changes

Faster dictation

• Decoder threads are bounded to 1 on GPU sessions, and the inference thread runs at user-interactive QoS. Benchmark under synthetic load: 30 s clip 12.8 s → 0.7 s, 90 s clip 17 s → 1.8 s.
• Transcribe during pauses (Settings › Behavior, on by default): completed speech is transcribed in the background at pauses; release transcribes only the tail and pastes once. Any missing or stale piece falls back to full-clip transcription. Release-to-text on 77–99 s clips: 0.7–3.3 s → 0.05–0.2 s.

Dictation accuracy

• Capture keeps 200 ms of audio after the physical release, so the last word isn't clipped. The release boundary itself stays exact.
• A final piece shorter than 3 s is transcribed together with the piece before it, so a lone word after a pause keeps its context ("…an amazing world" instead of "…an amazing").
• Piece joins drop false periods at mid-clause pause seams (after an article, "of", a hesitation, or a repeated word) and lowercase the continuing word, instead of splitting one sentence in two.
• Near-silent tails skip inference, and near-silent captures are discarded instead of hallucinating "thank you" (Inserting 'thank you' at the end of dictation sometimes #108).

Fixes

• Clipboard writes are marked transient, so clipboard managers skip dictations (Keyboard input simulation fails; clipboard insertion works kitlangton/Hex#223).
• ⌘V resolves through the layout's Command layer ("Dvorak – QWERTY ⌘", Insertion broken on "Dvorak — QWERTY ⌘" keyboard layout kitlangton/Hex#162).
• Release-microphone-while-idle no longer drops the first held press after a CoreAudio overload warning, and cold opens retry the selected device (First press-and-hold after idle records nothing: transient mic reopen failure suspends the in-flight hotkey gesture #130).
• Linux CI compiles again: the GGUF validator borrows the model path, and macOS-only history, model-deletion, and thread-test code no longer trips Linux dead-code lints.

Meetings (beta, opt-in)

• Settings toggle with Screen Recording health; a focus-free offer capsule for Zoom, Meet, Slack, Teams, and FaceTime calls; a menu-bar recording timer; a keep/stop prompt when the call app releases the mic. Recording never starts automatically.
• Rolling local transcription during the call, in overlapping chunks, with echo suppression by sentence plus a mic-vs-system energy check.
• Apple Intelligence (on-device) writes a provisional title while recording, then a final title and a summary of at most two sentences.
• Redesigned Meetings pane: search, day groups, notes first, a You/Them conversation view, live tail, rename, delete to Trash, and Markdown export to ~/Documents/Hex Meetings.
• hex mcp: a read-only stdio MCP server (list_meetings, search_transcripts, get_transcript, get_meeting_notes, Markdown resources). No writes, no network, no model calls. scripts/sync-meetings-mirror.sh mirrors another Mac's meetings for local MCP clients.

Scope

This branch combines perf/decoder-threads, perf/incremental-transcription, fix/triage-batch, test/dictation-corpus, and the meetings stack. Semantic transcript search, calendar titles, and the native macOS shell (docs/plans/native-macos-app.md) are follow-ups. The version bump to 2.2.0 (build 20200) lands separately on main as the release commit.

Verification

cargo fmt --check
cargo clippy --all-targets --all-features -- -D warnings
cargo test
git diff --check

• macOS cargo test: 584 passed, 13 ignored, plus the stdio MCP smoke test. Linux CI (test and nix) on this PR.
• Speed numbers come from a transcribe.cpp harness and an opt-in real-model test on the Mac Studio, not from live microphone capture. The release-boundary and short-final-piece results come from the dictation corpus harness (scripts/dictation-corpus.sh).
• Echo suppression was tuned on a real 17-minute speaker call: mic twins in the live view 105 → 2, with all 67 genuine You lines kept.
• Kit used a signed build of this branch on his MacBook Pro before approving the release. hex mcp was exercised over SSH from an OpenCode client: initialize 0.33 s, list_meetings 18 ms, search_transcripts 12 ms.
• Not individually verified: Dvorak – QWERTY ⌘ paste and clipboard-manager skipping.