Part of #99142

What Problem This Solves

room_event (ambient) turns have a strict contract: nothing reaches the source chat unless the agent explicitly calls message(action=send). Five paths violate it, so ambient groups receive unprompted bot output that no instruction can prevent: fast-mode auto notices ("💨Fast: auto-off(...)") bypass suppression on both the direct dispatch path and queued followups; verbose 🛠️ tool summaries deliver from queued room-event followups (while the direct path correctly suppresses them); followup-path compaction/preflight-failure/cross-channel-error notices deliver to the group; and queue-overflow summaries strip the room_event kind, turning a burst of ambient chatter into a normal turn with forceMessageTool — a direct invitation to respond to messages nobody addressed to the bot.

Why This Change Was Made

Each fix mirrors the existing correct gate rather than inventing new policy: the fast-mode dispatch bypass now requires ctx.InboundEventKind !== "room_event" exactly like the sibling verbose/forced bypasses; the followup paths gate on the queued run's currentInboundEventKind with logging consistent with the direct path's suppressed-notice logs; overflow summaries carry currentInboundEventKind: "room_event" only when every summarized item was a room event (mixed batches keep current behavior), tracked via an allRoomEvents flag on elision entries so the retry source preserves the kind.

User Impact

Ambient-mode group chats no longer receive fast-mode status lines, tool summaries, compaction warnings, or overflow-triggered replies the agent never chose to send. Non-room-event behavior is unchanged, including the existing "delivers fast auto progress in message-tool-only mode" contract.

Evidence

• Per-fix regression tests in dispatch-from-config.test.ts, followup-runner.test.ts, and queue.collect.test.ts: room_event variant suppressed, non-room-event variant unchanged; the overflow test's pinned undefined kind updated to the new all-room-events contract.
• node scripts/run-vitest.mjs over the three touched test files: 382 tests; both tsgo lanes; oxlint clean; repo autoreview clean; combined-tree validation with the two sibling PRs (both tsgo lanes + 16 test files + Linux Testbox gates).