feat(agents): add active turn indicators to Agents Menu

[wpfleger96]

Surface per-channel "Working in #channel" indicators in the Agents Menu so users can see at a glance which agents are actively working and where — without navigating to each channel individually.

Problem

The Agents Menu shows process state (running/stopped) and presence (online/offline), but not cognitive state. Users cannot tell whether an agent is in the middle of a turn without opening each channel and checking the activity panel.

Solution

Two-layer approach:

Backend (ACP harness): A TurnCompletionGuard scope guard in run_prompt_task() emits turn_completed on every exit path (success, error, timeout, cancel, panic) via its Drop impl. This is symmetric with the existing turn_started emit at the top of the function — every turn that starts is guaranteed to complete in the observer stream. The Drop payload is {} because the guard cannot know the actual outcome on the drop path.

Desktop (derived store + UI):

• activeAgentTurnsStore.ts — subscribes to the observer relay store and maintains a Map<agentPubkey, Map<turnId, ActiveTurn>> of active turns. Turns become active on turn_started and inactive on turn_completed / turn_error / agent_panic. A 90-second inactivity timeout prunes turns whose completion event was missed (checked every 5s).
• AgentStatusBadge gains a pulsing "Working" variant that takes priority over "running".
• ManagedAgentRow renders "Working in #channel-name" badges (max 3) below channel membership badges.
• All channel badges (both active "Working in" and static membership) are clickable, navigating to the respective channel on click.

Event ordering: composite watermark

The store dedups observer events with a composite (timestamp, seq) watermark per agent, reusing the compareObserverEvents comparator from observerRelayStore. Every event kind is gated on the watermark uniformly — an event is processed only if it is strictly newer than the last one seen.

• All event kinds are gated identically. Starts, activity, and evictions (turn_completed / turn_error / agent_panic) all skip equal-or-older events, so full-buffer replays are a complete no-op given the sorted-buffer invariant.
• Post-restart streams are handled for free: the harness resets seq to 1 on restart, but wall-clock timestamp keeps climbing, so the comparator accepts post-restart events without any seq === 1 special case.
• Why evictions are gated too. The harness emits turn_error / agent_panic with a null turnId, so the store falls back to deleting the first turn matching the channel. Processing such an eviction unconditionally on replay would re-run that channel-match fallback and delete the live turn whose start event is below the watermark. Gating evictions on the watermark closes that hole. Resurrection is not a risk: it would require reprocessing a stale start, which the watermark already blocks.

Exported API

Export	Purpose
useActiveAgentTurns(pubkey)	Hook: returns Set<channelId> where agent has active turns
useActiveAgentTurnsBridge(agents)	Bridge hook: syncs observer events into the store
getActiveChannelsForAgent(pubkey)	Snapshot accessor (for non-React consumers)
subscribeActiveAgentTurns(listener)	External store subscription
syncAgentTurnsFromEvents(pubkey, events)	Imperative sync (used by bridge)
resetActiveAgentTurnsStore()	Test utility

Design decisions

• Scope guard over manual emits: TurnCompletionGuard covers all exit paths structurally (including panics) without requiring each error handler to remember to emit.
• Observer-based over typing indicators: Typing indicators (kind:20002) have an 8s TTL that causes flicker. The observer stream provides explicit start/end signals and is already encrypted to the agent owner.
• 90s prune timeout: Safety net for missed completion events. Activity (acp_read / acp_write) refreshes lastActivityAt only for the matching turnId, so unrelated agent activity does not prevent expiry.
• MAX_TURNS_PER_AGENT = 4: Matches the ACP pool size. Oldest turn is evicted if exceeded.
• endTurn channelId fallback: If turn_completed arrives without a turnId (legacy path), falls back to removing the first turn matching the channelId.
• Clickable badges: Both "Working in #channel" and static channel membership badges navigate to the channel on click via goChannel(id). stopPropagation() prevents triggering the parent row expand.

Replaces #994 (which was opened from a fork); identical change set, head on block/buzz.

[wpfleger96]

Baseline — agents idle

Three agents: two running, one stopped. No active turns.

Single agent working

Paul is working in #general — shows "Working" badge and channel indicator.

Mixed states

Paul working in two channels, Duncan idle, Thufir stopped. Full state matrix.