feat(desktop): hold-to-dictate in the message composer

[chillerno1]

Closes #2708

Adds dictation to the message composer, inspired by the Claude Code implementation. The user can click the mic button in the composer toolbar, hold Space (~800 ms; a quick tap still types a space), or hold ⌃Space for an instant start. Words stream into the composer live from the on-device STT stack, releasing the held key commits the trailing phrase, and Enter while dictating sends the message once every in-flight word has arrived. Ships with a Voice & dictation settings section (Standard parakeet-en / Enhanced parakeet-v3 picker with download-on-select). Everything runs locally, so no audio leaves the machine.

How it works

• Composer (useDictation.ts): one hook owns a dictation session across all three triggers (mic button toggle, Space hold, ⌃Space). It anchors a text stream in the editor, replaces the live partial in place as re-decodes refine it, and commits on finals. Enter is intercepted at window capture during a session so the editor's own submit-on-Enter can't fire early.
• Rust (dictation.rs + huddle/stt.rs): reuses the huddle STT pipeline in a live-transcript mode. A self-pacing re-decode loop emits partials for the growing phrase buffer and a final at phrase end, over a strictly ordered channel.
• Enter-to-send is deterministic: stop_dictation pushes a zero-length sentinel after the flush silence, and the STT worker echoes it back as a Flushed event. The marker guarantees every transcript of the stopped session has been delivered, so the send fires on the marker rather than on heuristics, and rapid dictate/Enter/dictate cycles can't lose or splice messages.
• Model picker (huddle/models.rs): per-model download slots, resolve_dictation_model (preferred model if ready, else the startup model), and download-on-select with progress from the settings card.

Decode guards

While testing this I kept hitting cases where Parakeet's transducer lied on a re-decode of audio it previously got right: sentence fronts vanished, finals came back empty, and random noise turned into "Yeah." The three guards below came out of those sessions. Each one encodes a preference I settled on while dictating day to day rather than a rule from the model's docs, so they're the judgment calls that deserve the closest review (huddle/stt.rs, tests in huddle/stt_tests.rs; assertions use strings from my captured traces).

1. Collapse guard: every re-decode of a phrase hears a superset of the audio, so it can revise words but never honestly lose them. A decode with fewer words than the best partial is discarded, and a collapsed final commits the best partial instead.
2. Prefix stitch: after a mid-phrase pause the model sometimes drops everything before the pause while gaining tail words, so the word count still grows and the collapse guard passes it. A settled front of 4+ words never legitimately vanishes wholesale. When a re-decode disagrees on both of the phrase's first two words, its opening is anchored inside the sentence-so-far and the preserved front is spliced back on.
3. Lone-filler noise gate: a breath or hum that outlasts the VAD onset debounce decodes as a lone filler ("Yeah.", "Mm."). A phrase whose decodes never grow past a single filler is treated as noise and dropped end to end.

Two tradeoffs I made deliberately: I'd rather keep a phantom word the model hallucinated from noise than let an empty final erase a sentence I actually said, and I accept that dictating a bare standalone "Yeah." and nothing else gets dropped by the noise gate. Both felt right in practice; happy to revisit either if reviewers disagree.

Also in here

• Live-mode pipeline fixes: a wall-clock decode hold (prevents decode storms from overflowing the audio queue and splicing phrases), the audio queue deepened to 30 s, a ~96 ms VAD onset debounce, and empty finals propagating so stranded partials clear.
• The Space hold shortened to 800 ms; at 1.5 s people start talking before the mic opens.

Testing

• Unit tests for all three guards plus sticky punctuation in huddle/stt_tests.rs, using strings from captured live traces. #[ignore]d decode experiments document the measured model behavior that motivated the guards.
• Full suites green at 74683cac: cargo check/clippy clean, Rust 1634 passed, tsc + biome clean, frontend 3461 passed.
• I exercised it end to end while developing: long rambles, rapid dictate/Enter/dictate cycles, noise-only input (breath, hum, and clicks produce nothing), and model switching mid-session.

The 10 commits build up in review order: composer feature, settings and model picker, Enter-to-send, then the guards.

---

🤖 Generated with Claude Code