One event stops costing eight round-trips: scoped, coalesced room fetches

[udaychandra]

Why

The room's SSE handling was resilient but coarse: any spec-scoped event that wasn't a message triggered specReviews + listRuns + one reviewDetail per review, and a new message re-fetched a full page. Since presence landed, agent_tool_* and agent_presence events stream during live runs — so an open room fired the 2+N fetch storm on every tool call the agent made, each fetch its own SSH round-trip.

What

Route by event type instead of reloading on any match (spec mast-refetch-diet):

• Telemetry and presence events (agent_tool_*, agent_log_chunk, agent_presence, heartbeat) → zero fetches, zero timeline merges, zero renders in the room. Presence chips keep updating from the app-wide presence store, which was already wired.
• spec_message_posted → one messages fetch with the server's exclusive after cursor, now exposed through Gateway.listSpecMessages ({before, after, limit} options; the endpoint already supported it). If the announced message_id doesn't land in the merge — a cross-node message can sync in with an id older than the cursor — one full latest-page fetch recovers it. Echo suppression for your own posts is unchanged.
• review_* → that review's detail only, when the event carries the review id and the detail is loaded; the spec's review list otherwise. Verified in sail source: stage events carry no review_id today (only review_approved/finding_dismissed do), so the fallback is the common live path until sail adds it. review_iteration_started/review_completed are list-shape boundaries and refresh the list.
• Run-lifecycle events (spec_dispatched, agent_session_*, agent_failed, agent_cancelled, …) → listRuns for that spec only.
• spec_status_changed / snapshot / guardrail rows → render straight from the event payload, no fetch. The board's own reload path is untouched.
• Unrecognized spec-scoped events → the old conservative refresh, so a new server event type is never silently dropped.

Bursts coalesce per (kind, spec): microtask batching plus an in-flight dirty flag (the useBoard pattern extended to cover the in-flight window), so five mixed events cost one fetch per kind and an event landing mid-flight re-runs exactly once.

The reconnect gap-fill (#44) keeps its full reconcile — routing applies only to live events.

Tests

Behavior tests count gateway calls, per the spec's acceptance list:

• a remote spec_message_posted → exactly one messages call carrying after, zero review/run calls
• review_stage_passed with a review id → one review-detail call, nothing else; without the id → one list call
• 20 tool events + 2 presence transitions in an open room → zero fetches, zero new rows
• a burst of 5 mixed events → one coalesced fetch per kind
• unknown event type → conservative refresh retained

bun test (512 pass) and bun run typecheck green; no new dependencies; no sail changes.