chore(status): three debts the card was keeping to itself

[wine-fall]

Three open issues had no line in specs/STATUS.md ## Open, the section that
is the index for the backlog. Two were filed in this session; one had been
invisible since 2026-08-26.

• Interject a talk beat over a ducked song: the engine ducks, the director never asks [spec 03-02 §3.5] #163 (enhancement, eng) — the director never interjects a talk beat over
  a ducked song. The engine half has been complete since spec 03-02:
  Engine.play(voice) ducks any live music and pre-schedules the unduck at the
  clip's known end (src/engine.ts:346-353), and both the announce and a
  mid-song steer reply already ride it. Director.runVoice
  (src/director.ts:901) simply never asks — its music-segment race waits only
  on the song's end, the listener's next line, and a due switch's pick. The
  issue carries the staleness constraint on buffered beats and its dependency
  on the in-flight talk-context music-state fix.
• Music picks failing and running long: 3 dead probes, one 'nothing suitable', 82-192s picks in one session [spec 04 §3.3] #164 (watch, bug) — one real session on 2026-08-31 logged three dead
  stream probes, one pick that returned nothing suitable, and five picks
  running 82-192 s against the 40.2 s / 54.7 s baseline recorded in
  specs/spec04/04-no-dead-air.md §3.3. Source health is suspected and
  explicitly marked unverified; the numbers stay in the issue rather than the
  spec because the session was not run under the spec's measurement
  conditions.
• First-run onboarding keeps asking after /quit, and never says the answers were dropped #153 (bug, eng) — filed 2026-08-26, never given a line here.

Only the three pointers move. The issue bodies remain the record, per the card
rule in the section's own header and the murmur-issue skill. No code, spec,
or test is touched; STATUS.md is 62 lines, under the 85-line cap that
.github/scripts/check-pr.ts enforces.

AI coding brief

Original request — The listener asked for two issues to be recorded from
evidence already gathered in this session (the mid-song interjection feature,
and the music-pick failures seen in .dev/dev.log), then asked that the
resulting STATUS.md pointer lines not sit in the shared checkout waiting for
another PR to carry them, but ship on their own chore branch. The why: the
shared main checkout is used by several sessions at once, so an uncommitted
edit there is both invisible debt and a merge hazard.

Manual interventions — The listener supplied the evidence anchors directly
(the engine duck/unduck seam, the runVoice race, the log timestamps) and
set the design constraint that a buffered beat must be validated against the
live music state at air time. They also directed the follow-up: commit now
rather than letting the pointer lines ride a future PR, staging only
specs/STATUS.md.

Retro — Nothing was slow here. The one thing worth carrying forward: the
audit that surfaced #153's missing line came free from running
gh issue list --state open against the card, which is the murmur-issue sync
step — running it as a reflex whenever a new line is added catches this drift
before CI or a handoff does.

🤖 Generated with Claude Code