fix(slack): persist working pill across bridge restarts

[rogeriochaves]

Summary

When the bridge restarts (config-sync does this on every applied bundle, sometimes several times an hour while iterating), the in-memory active pill map is wiped. The 30s refresh loop is gated on that map being populated, so until the agent's next channel-root text post relights it, the pill stays dark. An agent that goes long on tools without an intermediate text post (common when debugging or planning) ends up looking dead in Slack for the rest of the turn.

Repro from earlier today: bridge restarted at 16:28 UTC, the dependabot-scout agent kept working but didn't emit channel-root text until 18:31 (it was "Manifesting" on a long Bash sequence the entire time). Channel showed no progress and no pill for ~2 hours.

Fix

Persist active to ~/.kanban-code/active-pills/<slug> (same atomic write-then-rename pattern as thread-root), restore on bridge startup. Two guards on restore:

1. Skip pills older than 10 min — Slack's own idle TTL would have cleared the visual pill by then anyway, and an agent that's been silent that long has likely finished its turn. Re-lighting would falsely advertise active work.
2. Re-light immediately on restore rather than waiting for the next refresh tick — gap between bridge start and visible "is working…" becomes seconds instead of a minute.

Persist on every set / refresh / clear so the on-disk state stays in lockstep with the in-memory map.

Test plan

• 6 new unit tests for active-pill: round-trip, missing file, corrupted file, partial record, clear, idempotent clear. 242/242 passing.
• Build clean.
• After merge + box pulls main, trigger an agent (e.g. force-sync the bundle, which restarts the bridge), confirm the pill survives the restart on a long-running tool sequence.