Voice: a dot-matrix glyph on every surface, phone dictation through the host

[gaelcado]

Summary

One dot-matrix voice glyph for every surface, on desktop and iPhone, replacing the orb, plus dictation from the phone through the chat's host.

• Shared glyph engine (zeron-orb::glyph): a field (the call's disc) and a beam (a strip) decide every dot's level, colour and size from the two voices' bands and the call's mode, on both platforms. Light is a halftone: dots swell and touch as they light. Each dot takes one of the working glyph's three inks (GSPIN_ROW_TINTS), scattered at random and never blended; the user's voice takes the accent. Modes pass into each other through a layered passage so a change mid-fade joins the blend. The old orb engine, its presets and the thinking-orbs chain (and its licence) are gone.
• Desktop: the stage orb in the composer's own disc, graded from a hot heart to small dots at the rim, lit under the pointer; a call tray above every conversation's composer on the queue's surface, its beam the dictation track's twin; the beam in miniature in the sidebar footer. The user's voice is read from the microphone's own spectrum (zeron-voice::spectrum), gated while Codex plays so playback does not come back as the user. Continuous loud speech no longer pins the glyph: the loudness scale runs to full scale and the heart beats with the syllables.
• iOS: GlyphView paints the engine's frames through the mobile FFI for the stage, the tab accessory and the bar item; Live Activity assets follow. The stage anchors the orb under the host name and gives the words the whole band to the controls.
• Transcript (both platforms): every turn of the call is kept and a turn's whole text with it, centred in its band while it fits and scrolling once it does not, the newest in view. The transport's 32 KB per-message guard is unchanged; the caption's own ceiling is 1 MiB per turn and keeps the beginning.
• Captions: Codex's item/started carries the role before any delta, so a turn is attributed as it begins; one still unnamed after a short grace is the user's, and the provider's word wins.
• Dictation from the phone: no model on the phone. It captures, resamples to 16 kHz PCM and streams over the device relay to the chat's host, whose own Parakeet worker (zeron-voice, fed from the network) writes it down and sends partials and the final back. Tap to start, tap the stop to transcribe. The host advertises dictation-host-v1; the RPC surface is StartDictationV1, FeedDictationV1, FinishDictationV1, CancelDictationV1 and the OwnDictationV1 stream. One dictation at a time per host, shared with its own composer; an unfinished one is dropped after 90 s.

Look closely at

• Compatibility: VoiceEvent::Partial gains an optional role (#[serde(default)]); dictation is gated on the dictation-host-v1 capability, and a phone talking to an older host gets a toast instead of a failure. No persisted keys or RPC names change.
• crates/voice-session keeps every caption turn now (was 64) and no longer restarts a turn past 32 KB. The mobile state carries all turns on every update; fine for hundreds of turns, unmeasured beyond.
• The desktop reads the microphone a second time for the glyph (zeron-voice::spectrum::Listener), levels only, only while a call is live and unmuted; ZERON_VOICE_SPECTRUM=0 turns it off.
• New dependencies: zeron-mobile gains base64 and uuid (already in the workspace) for the dictation FFI; zeron-engine depends on zeron-voice.

Device deletion (a deleteDevice mutate op) was developed alongside and is split out for its own PR.

Test plan

• cargo test --locked -p zeron-ui --lib -- --test-threads=1 (1776 passed)
• cargo test --locked -p zeron-engine --lib
• cargo test --locked -p zeron-harness: 319 passed; two timing-bounded tests (executable::tests::version_probe_bounds_hangs_and_rejects_nonzero, opencode::tests::stalled_prompt_post_has_a_bounded_timeout) failed once while the other suites ran alongside and pass in isolation; code untouched by this PR
• cargo test --locked -p zeron-sync --lib
• cargo test --locked -p zeron-preview: not completed here; its tests/leak integration test sat idle for 15 minutes on this machine and was killed (crate untouched by this PR)
• cargo test --locked -p zeron-orb -p zeron-mobile -p zeron-voice -p zeron-voice-session -p zeron-proto -p zeron-doc -p zeron-rpc
• cargo test --locked -p zeron-engine --test workspace_sync
• cargo clippy --locked --all-targets on the touched crates: no new warnings
• New tests: glyph engine (pointer touch; a mode arriving mid-passage joins the blend without a jump), caption (every turn kept; a UTF-8 turn past 32 KB keeps its first and last words), workspace capability fixture, iOS DictationDraftTests
• Manual, macOS dev build: stage, call tray, dictation track, footer in dark and light, frosted; stage transcript centred with 1–4 turns then scrolling with a top fade; pointer over the orb
• Manual, iPhone 13 Pro dev build and simulator: stage in dark and light, transcript centred then scrolling, tab accessory and bar glyph, tap-to-dictate UI over a scripted transcriber
• Not done: a live phone-to-host dictation against a host with the model; a live call judged for fast-speech reactivity (the preview feeds a synthetic wave); Live Activity on a real lock screen

Screenshots

Desktop stage, call strip and dictation track in dark and light; iPhone stage in dark and light. Attached below.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}