(CODUX-219) Implement sentence-aware speech chunking

[numnx]

🧩 Task: Implement sentence-aware speech chunking

Automated task PR opened by Code UX for T03.

📋 Summary

Sprint: Sprint 219 · Fix chat continuations and TTS playback · Task: T03

Keep agent-owned Project Manager continuations bound to their originating chat thread without cancellation cascades or duplicate turns, and deliver long Chat replies through a fast, sentence-chunked, ordered TTS pipeline.

🤖 Model & Provider

Provider	Codex CLI
Model	gpt-5.6-sol

⏱️ Timing

Started	2026-07-13 17:04:24 UTC
Finished	2026-07-13 17:11:56 UTC
Duration	7m 32s

📊 Token Usage

Metric	Value
Input tokens	92,610
Cached input tokens	1,544,960
Output tokens	16,542
Total tokens	1,654,112
Invocations	1
Tool calls	28
Active time	6m 55s

Included usage estimate (subscription/local login): $1.73

✅ QA Review

⏳ QA review runs after this PR is opened. Check the task card in the dashboard for the outcome once it lands.

📝 Full Task Prompt

## Objective
Create a deterministic speech-text chunker that lets long Project Manager replies begin with a complete sentence quickly while preserving every speakable word in order and respecting the backend synthesis limit.

## Scope
- `dashboard/src/v2/lib/speech-playback.ts`
- `dashboard/src/v2/lib/__tests__/speech-playback.test.ts` (create)
- Existing Markdown/widget speech filtering in `speechTextFromMarkdown`

## Implementation Requirements
1. Split normalized spoken text at sentence boundaries, including punctuation and common closing quotes/brackets, while retaining punctuation needed for natural speech.
2. Group or fallback-split chunks so every request stays within `MAX_SPEECH_PLAYBACK_CHARS`; handle long unpunctuated text, a single oversized sentence, repeated whitespace, empty Markdown, code fences, and dashboard-only widget fences without dropping or duplicating content.
3. Add tests proving chunk order, exact normalized-content reconstruction, fast first-sentence availability, punctuation handling, maximum-length enforcement, and preservation of the existing code/widget suppression behavior.

## Constraints
- Keep the 8,000-character request ceiling compatible with `POST /api/speech/synthesis`.
- Do not send Markdown decoration, ordinary fenced code, or `codeux:*` widget payloads to a speech provider.
- Keep the helper pure and independent of browser audio, network calls, React/Preact state, or external packages.

## Verification
- Run `pnpm exec vitest run dashboard/src/v2/lib/__tests__/speech-playback.test.ts`.
- Run `pnpm run typecheck:dashboard`.
- Confirm all chunks are non-empty, ordered, bounded, and reconstruct the filtered spoken text exactly.

🌿 Branch Info

Base: feature/CODUX-219-fix-chat-continuations-and-tts-playback
Head: task/feature-codux-21-t03-codex-32e6d7af-mrjh2mcn

---

🤖 Generated by Code UX