Live voice (experimental), voice orchestration, P0 fixes, and composer QoL

[ojowwalker77]

Rebased onto main (v0.12.0). Everything here is a single commit — the branch was all working-tree work before this PR.

Live voice (experimental, Codex)

• WebRTC microphone input + streamed assistant audio playback (remote-track <audio> sink).
• Realtime transcripts and Thinking Orbs voice states (connecting / listening / speaking / muted).
• Full-screen voice-focus surface that owns the transcript pane (sidebar/header/composer chrome preserved).
• Mic/audio frames kept out of the durable orchestration event log; only control state + final transcripts persist. SDP preserved as opaque protocol data (trailing CRLF intact).
• Capability-gated: the voice affordance only appears on Codex threads.

Voice as orchestrator

• TeaCode MCP thread controls — threads_list / threads_create / threads_send / threads_read — so voice can spawn separate-context threads and drive other providers.

P0 fixes

• No implicit browser auth on startup. Cursor and Grok are gated behind explicit connect; Codex / OpenCode / Kilo / Pi / Claude audited clean. XAI_API_KEY still works with no browser.
• New Thread exposes the real model picker (the centered tray previously omitted it).
• Settings: default thread workspace behavior — new worktree / current branch / pinned base branch.

Composer QoL

• Per-provider quota ring in the composer underbar: remaining-% + a ring that fills for what's left and recolors green→red as it drains; click for the full usage popover. Reuses the existing tone/usage logic; hidden when a provider has no usage data.
• Do-things ⌘K palette: switch model, switch provider, and spawn a new thread pre-set to a chosen provider — all from the keyboard.

Deferred by design (not bugs)

• ⌘K branch-switch (needs cwd derivation that only lives in ChatView today — TODO seam left).
• ⌘K voice toggle (its hook is bound to ChatView's lifecycle).
• Error surfacing: traced end-to-end and already correct — server preserves detail/cause, the thread banner shows the raw message, toasts use err.message. The one real offender (voice masking a generic stderr line) was already fixed.

Verification

• bun fmt / bun lint (0 errors) / bun typecheck (8/8 packages) green on the integrated tree.
• Affected web tests: 119 passing (voice, realtime audio, quota ring, palette commands, appSettings, threadBootstrap, wsNativeApi). Server auth-gating tests: 106 passing.

⚠️ Live voice depends on OpenAI's experimental realtime backend and a separate voice allowance — availability is upstream-gated.