Chunk speak text and disable TTS keepalive deadline for long --url input

[alexkroman]

Problem

assembly speak --url https://arxiv.org/pdf/... (a long PDF/article) failed with:

Error: TTS session failed: sent 1011 (internal error) keepalive ping timeout; no close frame received

The "no close frame received" means the client (websockets) gave up on a still-alive socket, not that the server crashed.

Root cause

Two independent contributors, both fixed here:

1. The whole document was sent in a single PocketTTS Generate frame. PocketTTS is a streaming model meant to be fed incrementally; a whole paper stalls the server long enough that it stops answering the websocket keepalive ping, and the client closes the socket with code 1011.

2. websockets' default 20s keepalive pong deadline is too aggressive for this workload. A server slow to emit the first Audio frame under load gets killed before producing anything.

Fix

• session.synthesize_chunked splits the text into sentence-aligned chunks packed to a safe char budget and synthesizes one connection per chunk — the same one-sentence-per-connection pattern agent-cascade already uses. An over-long, terminator-less PDF blob is hard-sliced so no single Generate can blow past the server's input ceiling. Bonus: audio now starts on the first chunk instead of after the whole document.
• ping_timeout=None on the TTS socket disables the redundant pong deadline. _RECV_TIMEOUT_SECONDS (60s) is already the per-frame liveness authority, so a slow-but-alive server is no longer killed; a genuinely dead connection still fails cleanly.

Notes

• New aai_cli/tts/text.py holds the pure split_sentences/chunk_text helpers (Rich-free, unit-tested).
• The oversized test_tts_session.py was split along its natural seam (single-synthesis vs dialogue) via a shared tests/_tts_session_helpers.py to stay under the 500-line file gate.
• Full scripts/check.sh gate passes (coverage, 100% patch coverage, mutation gate, build).

🤖 Generated with Claude Code