fix(director): the beat speaks from the real program, not an imagined one [spec 04]

[wine-fall]

Fixes the self-initiated talk beat fabricating music state. Evidence (.dev/dev.log 2026-08-31): in the no-song stretch after a 14:06 music.end, beats claimed a track was still playing and ending; during the 14:23–14:28 pick-failure window (dead probe, "nothing suitable found") the beats invented an entire broadcast — a song chosen, put on, heard to its end, the listener's kettle heard boiling — and called the afternoon over at 14:28.

Root cause: the talk prompt carried no music facts and no clock — only the six-hour scene bucket — so the model filled the silence in. Beats also air minutes after generation via the spec-04 look-ahead buffer.

Related spec: specs/spec04/04-no-dead-air.md (§3.4 context enrichment — bugfix, spec unchanged per convention).

What changed

• ContextPack gains time (real local clock, ICU-free formatClock) and music — a discriminated MusicState: playing (the on-air label), quiet (nothing playing, optionally the last track), picking (search in flight), pickFailed (latest search empty). Derived in the Director from the segment, the pick slot, the search outcome, and the last aired track; a wired radio always states a music fact.
• The talk builders render both plus anti-fabrication red lines: the system announces tracks (the startTrack announce path, unchanged); the beat never narrates picking/starting a song, never claims listener-side sounds, never narrates time passing.
• Peer-review hardening: the respond/steer prompts get the same clock + music facts (without the beat-shaped red lines); a dropped pick resolving late is epoch-guarded out of pickFailed; a forced MURMUR_SCENE drops the clock line rather than contradict it.

Unchanged by design: mood-driven pick selection (spec 04), gapSeconds, the announce path.

Peer review (codex gpt-5.6-sol, xhigh): 4 findings, 4 applied, 0 dismissed.

Gates: 1012 vitest green, tsc, oxlint. By-ear: whether the beats actually stop fabricating on a real run is a stochastic outcome — worth an ear during the next normal listening session, no dedicated pass owed.

AI coding brief

Original request: Fix the bug of self-initiated talk beats fabricating music state (songs announced that never played, imagined listener-side sounds, time narrated faster than real), with evidence and the fix plan pre-settled in-session: ground the ContextPack with real music status + real clock, and add prompt red lines.

Manual interventions: None during the build — the invocation carried the full investigation (log evidence, root cause, fix design, constraints), so the session went straight to test-first implementation.

Retro: Front-loading the diagnosis into the /murmur-ship brief (evidence timestamps, named code seams, settled design, explicit non-goals) let the build skip re-investigation entirely; the only additions came from the codex closing review, which the brief correctly left open.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MSvKfUaQzE95K77oqaBzDU