Brain into Rooms: convergence, persisted history, reply-loop fix, manual halt

[Wirasm]

Follow-ups from the deep-dive review of the recent channel → Room rename. Five threads, one PR.

1. Agent prompts say "room" (#1)

.pi/agents/{orchestrator,worker,reviewer}.md were still describing a "channel" — the data agents actually read. Updated to "room" / "participant(s)" to match the primitive.

2. Brain converged onto the real Room primitive (#2)

The brain talked over a separate CommsBus while the cockpit used the real Room primitive — two things both called "room". Now the brain drives the same surface as the human:

• brain.ts gains open_room / post_to_room backed by roomManager; RoomManager gains postAs / messages.
• Deleted comms-bus.ts and the superseded rooms-demo.ts (the real multi-agent room demo is kild room --participants a,b,c).
• Agent dispatch stays on Flue (run_agent_in_worktree) — real participants are worker subprocesses that only bootstrap inside the engine, so standalone workflows can't spawn them.

3. Room log persisted as read-only history (#3)

Rooms were in-memory only and vanished on every --watch reload. RoomRegistry now write-throughs each room's log to $KILD_HOME/rooms/<id>.json and loads past rooms into a read-only archive on start. New GET /api/rooms/archive; the cockpit renders archived rooms as a read-only transcript (disabled composer). Participants stay dead — it's a record, not a resumable session.

4. Fixed the agent reply-loop (#4, found live)

Two agents ping-ponged "standing by" / "(no response needed)" forever and burned tokens. Root cause: the implicit reply auto-posted an agent's turn-final narration as a delivered turn — so "I'll stay quiet" became a message that prompted the other agent, contradicting the prompts ("no one sees your normal output — only what you post"). Fix: an implicit reply now broadcasts but never delivers a turn; prompting another agent requires an explicit post_message @mention. Two router tests lock it in.

5. Manual circuit breaker (halt button)

A Topbar ⏹ halt button (room_halt) stops a room's agent sessions but keeps it read-only — the operator can trip it to stop a runaway or off-track room without tearing it down (vs ✕ which closes it).

Verification

• engine: typecheck ✅, lint ✅, 44 tests ✅ (+4), compile ✅
• app: svelte-check ✅ (0 errors)
• Ran engine + cockpit locally; health + archive endpoint verified.

Notes

• room-types.ts and a few files span more than one thread, so this is one commit rather than split.
• Loop control beyond the root-cause fix is now the manual halt button (no automatic turn cap), by request.