fix(director): air talk instead of blocking on an unready music pick [spec 04]

[wine-fall]

Problem

A real listening run (.dev/dev.log) opened with ~63s of dead air before the
first song. A talk beat was already synthesized and buffered, but when music was
due the branch entered _play_music_segment → _take_pick and blocked ~56s
awaiting a not-yet-resolved prefetched pick instead of airing the warm buffered
talk. The dev-log seam shows it plainly:

music.pick  found=True prefetched=True elapsed_s=56.52

prefetched=True with a 56s elapsed means the prefetch was in flight but not
done, and the music branch awaited it — blocking the whole loop.

Root cause: the music branch awaited an in-flight prefetch, blocking the loop
even when talk was ready.

Fix

Guard Director._play_music_segment: when the prefetched pick is still
resolving (in flight, not .done()), return to talk and re-attempt music at
the next boundary. The pick keeps resolving in the background — the slot is not
cleared, so no duplicate prefetch fires — and the depth-2 talk look-ahead
(spec 04 §3.3) covers the search adaptively, however long it takes, with no
dead air and no hardcoded "talk for N minutes" duration.

The cold path (nothing prefetched — the first-ever segment) and the resolved-pick
path are byte-for-byte unchanged; only an in-flight pick now defers instead
of blocking.

Tests

• New regression test_music_does_not_block_on_an_unresolved_pick_airs_talk: a
  fake TrackSource whose pick never resolves now yields continuous talk and
  the run completes under max_segments. Pre-fix the music branch awaited the
  pick and the run hung (the test times out on old code).
• Full unit suite: 250 passed. Every existing music-first test is preserved
  (they exercise the untouched cold path).

Spec

Aligned specs/spec04/04-no-dead-air.md §3.1 Consume bullet (no longer "finish
what's left" — a resolving pick now defers to talk) and added acceptance
criterion #4 (no-block / dead-air).

Implements specs/spec04/04-no-dead-air.md

Peer review

Peer review (codex gpt-5.5): 1 finding, 0 applied, 1 dismissed. The finding — a
resolved pick held across several talk boundaries under erratic
RandomCadence/BrainCadence grows stale and blocks a fresher prefetch — is a
spec-accepted trade-off (§3.1 "Staleness (accepted)" + §6 open question on
mood-staleness tolerance) and the intended single-slot "one pick ahead" design
(re-prefetching every talk would waste the expensive yt-dlp resolves slice 1
exists to avoid). Not a regression: the change adds at most one deferral boundary
to an already-accepted staleness window, and music airs whenever cadence next
chooses it.

AI coding brief

Original request: handoff scratch/handoff-spec04-followups.md Task 1 — a
real listening run (.dev/dev.log) showed ~63s dead air before the first song
because the music branch blocked ~56s awaiting a not-yet-ready prefetched pick
instead of airing the already-buffered warm talk. Why: cold-start
responsiveness / no-dead-air (spec 04).

Manual interventions: none — the fix followed the handoff's diagnosis
directly.

Retro: the handoff pre-diagnosed the root cause with the exact dev.log line,
which made this a one-guard fix. Precise seam-level evidence (the
prefetched=True elapsed_s=56.52 line) is what turned a vague "dead air" symptom
into a targeted change — capture the deterministic seam in the bug report and the
fix writes itself.

🤖 Generated with Claude Code