perf: polish chat streaming and queue UX

[0xSolarPunk]

What changed

This brings a few Codex-style UI/perf ideas into Garcon chat without changing the server/API contract.

• Streaming output: batches agent output received in the same WebSocket drain so chat message state is updated once per drain instead of once per tiny streamed chunk.
• Composer drafts: debounces draft persistence so typing does less localStorage work, while still flushing pending drafts on teardown, submit, and chat changes.
• Queue UI: replaces the plain queued-message list with a compact preview panel: count badge, up to 3 visible entries, long-text truncation, hidden-count text, and Queued inputs / Sending labels.
• Safer queue controls: sending entries are visible but no longer show the remove button, because removing them would not cancel the already-dispatched run.

Review fixes

• Flushed batched streamed output before same-drain stop/error transcript messages, preserving transcript order.
• Added regression coverage for same-drain output ordering.
• Added regression coverage for display-only sending queue entries.

Side effects / risk

• Draft saves now happen after a short debounce instead of every keystroke; a hard tab crash inside that small window could lose the latest keystroke, but normal cleanup flushes it.
• Streaming UI updates happen once per WS drain; this is the intended perf win and still flushes before lifecycle/error messages.
• No backend, API, or WS contract changes.

Screenshots

Desktop: https://files.catbox.moe/aukqhf.png
Mobile: https://files.catbox.moe/b0jrmt.png

Validation

• bun run check
• bun run test
• timeout 30s bun run start --port 0
• GitHub checks green: server-tests, web-tests, build-exe