feat(director): talk look-ahead survives music, depth-2 buffer [spec 04]

[wine-fall]

Implements specs/spec04/04-no-dead-air.md §3.3 — the talk look-ahead now survives music.

Problem

The talk look-ahead buffer was discarded when a song went on air and never refilled during playback. So the music→talk boundary always hit a cold next_talks + synth — dead air after every song. The music side already prefetched during talk (slice 1); the talk side had no symmetric prefetch during music.

What changed

A depth-2 talk look-ahead buffer (_TALK_LOOKAHEAD) that mirrors the music slot:

• Survives music, refilled when drained. A song no longer discards the buffer — a song is the ideal window to prepare the next talk. _prefetch_talk refills when the buffer is empty (one coherent next_talks(N) batch, beats synthesized in parallel), fired after an aired beat is recorded and at a music segment's start, so the refill's Brain+synth overlap the song and the post-song talk airs warm.
• Coherent batches. Refilling only-when-empty and only-after-record means each batch regenerates over context that already includes every prior aired beat (the real brain is stateless) — no duplication against beats still buffered.
• User priority preserved. A talkback Steer cancels an in-flight refill without awaiting it, so a reply is never blocked on a background prefetch's teardown (spec 01 §3.3); the refill settles on the next refill or at shutdown.
• Resilience. Bounded retry (_LOOKAHEAD_ATTEMPTS) on the refill's next_talks and on synth, degrading (lose the look-ahead / the one beat, never the radio); every stage logs to the dev log.

Tests

New fakes for transient next_talks / synth failures. New unit tests: survives-a-song (buffered beat airs post-song, no cold regen), depth-2 across two music completions (both post-song talks warm), Brain-retry, synth-retry. Updated the tests whose contract changed (song no longer discards; steer-resume asserts the invariant not an exact beat number). pytest: 248 passed. pyright + source-language gates green.

Verified at the deterministic seam (fakes) per the acceptance criteria; the end-to-end first-music latency re-measurement (the ~76s that motivated spec 04) is tracked as still-open in specs/STATUS.md.

Peer review

codex (gpt-5.5, xhigh) vs origin/main: 2 findings, 2 applied, 0 dismissed — (1) cold/refill using stale context could duplicate buffered beats → fixed with refill-when-empty + refill-after-record; (2) reply blocked on a stale refill's cancellation → fixed with cancel-not-await on the steer path.

AI coding brief

• Original request: While reviewing spec 04, the user observed that at startup two TTS beats + a music pick are prepared, but during music playback nothing regenerates the next talk — so the music→talk boundary is cold. They asked to build the missing symmetric prefetch: a talk slot, capacity 2, that always refills from current context when drained ("just like the music-find logic"). The "why": remove the dead air after songs.
• Manual interventions: The user set the acceptance bar explicitly — buffer capacity 2 (cover the next two music completions), parallelize where possible, and retry + logging on the important stages. codex review then redirected the refill design (refill-when-empty + after-record) to avoid stale-context duplication and to keep replies off the refill-teardown path.
• Retro: Prompt faster next time by stating the capacity/parallel/retry/log constraints up front (the user did) — they shaped the design cleanly. The one avoidable detour was committing onto a stale already-merged branch; branching fresh from main per feature would have skipped the cherry-pick.

🤖 Generated with Claude Code