fix(desktop): sync derived active-turn liveness in channel activity path

[tellaho]

Overview

Category: fix
User Impact: The working/liveness indicator in a channel's agent Activity panel now clears as soon as the agent's turn completes, instead of continuing to spin after the work is done.

Problem: In the channel Activity panel, a raw turn_completed observer event could be visible in the feed while the liveness indicator kept spinning. The channel screen mounted the observer bridge (which feeds raw ACP frames into the activity panel) but never mounted the active-turns bridge, so the derived liveness store went stale — the exact "raw event arrived but the non-raw feed didn't update" mismatch. Nearby surfaces (profile panel, agents page, sidebar) already mount both bridges and were unaffected.

Solution: Mount useActiveAgentTurnsBridge right next to the existing observer bridge in ChannelScreen, on the same agent list, so channel-activity derived liveness consumes the same observer events the raw rail does. The bridge's sync loop is extracted into an exported syncActiveAgentTurnsFromObserver so a regression test can drive the real observer-ingestion path without a React renderer. Also wires resetActiveAgentTurnsStore() into the workspace-switch reset, closing a singleton-leak gap per the workspace-switching contract in AGENTS.md.

File changes

desktop/src/features/channels/ui/ChannelScreen.tsx
Mount useActiveAgentTurnsBridge(observerBridgeAgents) alongside the existing useManagedAgentObserverBridge so the derived active-turns store stays in sync with raw observer frames in the channel activity path. The store dedupes per-agent via a watermark, so double-mounting with other surfaces is a no-op.

desktop/src/features/agents/activeAgentTurnsStore.ts
Extract the bridge's sync loop into exported syncActiveAgentTurnsFromObserver(agents); the hook now delegates to it. No behavior change — enables the regression to exercise the exact production path. Also resetActiveAgentTurnsStore() is now invoked on workspace switch.

desktop/src/features/agents/activeAgentTurnsStore.test.mjs
New regression "observer → active-turns bridge sync": injects turn_started then turn_completed through the real observer ingestion path and asserts the raw feed retains the completion event while derived liveness clears. Plus a case asserting non-running/deployed agents don't populate the store.

desktop/src/features/workspaces/useWorkspaceInit.ts
Add resetActiveAgentTurnsStore() to resetWorkspaceState() — the active-turns store is a module-level singleton holding workspace-scoped state and previously leaked across workspace switches.

desktop/src/features/agents/ui/ManagedAgentSessionPanel.tsx
Net-zero change from the temporary debug harness used during diagnosis (added and fully removed within this branch); only minor prop-ordering churn remains.

Reproduction Steps

1. Run the desktop app with a local managed agent and open a channel the agent is in.
2. Mention the agent so it starts a turn, then open its Activity panel from the channel (bot activity bar or member menu).
3. Watch the turn finish: previously the liveness indicator kept spinning after turn_completed appeared in the raw feed; now it clears promptly.
4. Automated: cd desktop && pnpm test — see the "observer → active-turns bridge sync" regression in activeAgentTurnsStore.test.mjs (1551/1551 pass).

No visual/markup changes — this fixes the state plumbing the liveness UI depends on.

[wpfleger96]

🟢 Nit: the comment removed just above this line was load-bearing context.

The deleted two lines explained why lastMessageAt is not used as an activeReadAt fallback — it's reply-inclusive and would clear unread thread/sidebar state before a real top-level position is known. The logic is unchanged and correct, but that constraint isn't obvious from the surrounding code. Suggest restoring the comment so the null fallback stays self-documenting.

[tellaho]

Updated by Ned 🤖 in bb069eb.