feat(director): thread a time-of-day scene into the talk prompt [spec 04]

[wine-fall]

The radio's scene was permanently "night" (the persona seed was night-flavored).
This threads the local time of day into the host's talk so the phrasing
varies by hour.

• scene.scene_for(now) is a pure, clock-free function mapping a datetime to a
  coarse bucket — morning 05:00–11:59, afternoon 12:00–17:59, evening
  18:00–22:59, late-night 23:00–04:59 (midnight wrap). The Director supplies the
  real datetime.now() where it builds the ContextPack; scene_for is the
  tested seam.
• ContextPack gains an optional scene: str | None (additive). The
  self-initiated talk prompts append a per-scene mood cue keyed by ctx.scene
  (a None / unmapped scene appends nothing). The persona seed is generalized
  off its hard late-night framing so the cue sets the mood.
• Deterministic parts unit-tested: bucket boundaries + midnight wrap with an
  injected clock (never datetime.now()), and prompt assembly (cue threaded
  in / omitted when unset). Whether the host's phrasing actually reads as
  morning vs. late-night is an eval / by-ear item, not a unit assertion.

Implements specs/spec04/04-no-dead-air.md (new §3.4).

By-ear / eval item owed (not self-declared)

The deterministic seam is proven on fakes (scene reaches the prompt). The
stochastic half — does the real Claude host actually sound like morning vs.
late-night — is owed as a listening pass: run make dev (or make dev-fishaudio)
at two different local hours (or override the clock) and confirm the opening
beats' mood shifts with the scene. Not asserted here on purpose (DESIGN §10.3).

AI coding brief

• Original request: The radio's scene is always "night." Read the local
  machine time, derive a scene (morning / afternoon / evening / late-night), and
  thread it into the Brain's context so the host's talk varies by time of day.
  The deterministic scene derivation must be unit-tested with an injected clock;
  the phrasing quality is eval-track, not a brittle assert on model text.
• Manual interventions: The design fork was surfaced as a clarify-gate before
  coding. The user chose (1) spec home = fold into spec04 — chosen with the
  caveat visible that spec04 is about latency, not context content, so §3.4 is
  framed as an adjacent add and the proper spec-07 time home is still flagged as
  owed; and (2) 4 buckets with the hour cuts above.
• Retro: Surfacing the spec-home + bucketing fork up front (with a recommended
  option + trade-offs per choice) avoided a mid-build rework. For prompt-threading
  features, stating "deterministic (bucketing + prompt assembly) → unit; phrasing
  quality → eval/by-ear" in the plan keeps the test line from drifting into a
  brittle assert on model output.

🤖 Generated with Claude Code