fix(bias): don't let an unreachable bulb spin the edge-latch loop

[cnewkirk]

Problem (observed live 2026-07-25)

A housekeeper unplugged the Couch lightstrip — a TV-bias light. Its write
failed on every tick, and the daemon only latched a committed bias edge
(_bias_last_applied_on) when all per-light writes succeeded. So the edge
never latched, the 2 s control-file poll read every cycle as a fresh TV flip,
and the daemon re-fired a 2.0 s edge fade to all 7 viewing lights every ~2 s
for hours (293 edge fades in 3 minutes). Symptom: the whole viewing set
visibly dimmed/brightened continuously at dawn.

The don't latch a failed edge rule was built for transient bridge
rejections (e.g. "command queue is full") that clear on retry — it had no
concept of a permanently unreachable device.

Fix

Classify write failures instead of collapsing them to a bool:

• BridgeError gains an unreachable flag; HueClient._parse sets it when the
  v2 errors array reports a device has "communication issues" (the
  unplugged-bulb signal) — the string match lives in exactly one place.
• _write_light returns _BiasWrite.OK / FAILED / UNREACHABLE. The edge
  latches once every write has landed or hit an unreachable device; only
  FAILED (a transient reject worth retrying) holds the latch open.

A dead viewing light now degrades to "that bulb is dark" — the others behave and
nothing cascades. Transient-retry behaviour is unchanged.

Tests

• test_unreachable_light_does_not_stall_edge_latch — edge latches despite a
  dead bulb; the next apply is a no-op (no storm); the reachable light still
  held.
• test_parse_flags_communication_issue_as_unreachable /
  test_parse_transient_error_is_not_unreachable — parser classification.
• Existing test_failed_edge_writes_retry_on_next_apply still green (transient
  retry preserved). Full suite: 363 passed, mypy clean.

Deploy status

Already built and running live on the NAS hue-circadian container (recreated
from the identical prior spec; old container parked as hue-circadian-old for
rollback). Verified healthy, day-drive normal, bias edge-fires down to ~1 per
300 log lines.

🤖 Generated with Claude Code