[TrueFurina]

Summary

This PR fixes the active-turn flag leak reported in issue #2456, where a Dashboard thread could stay urn_active=true forever after a dropped WebSocket handler task, a client disconnect mid-stream, or the process being interrupted. Once stuck, the composer stays disabled ("???") and the user cannot type a new message.

Root cause

WebSocketHub._active_turns was a plain set. mark_turn_idle only runs inside the inally block of WebSocketChannel.handle_inbound. When the coroutine is dropped (connection loss, process crash, task cancellation), that inally is unreachable - the flag leaks for the lifetime of the process.

Fix

1. _active_turns is now dict[str, float] mapping hread_id -> time.monotonic() of the last pushed frame.
2. push_to_thread auto-refreshes the TTL via a new �ump_turn_activity helper, so any thread that still receives chunks stays "warm".
3. New evict_stale_turns(max_quiet_seconds=30.0) drops entries that have not been bumped for more than 30 s, and logs which thread_ids were pruned. It is called lazily from is_turn_active, so reconnecting clients (or the ping machinery) heal the leak without any scheduler.
4. mark_turn_active / mark_turn_idle keep their public signatures; only the backing store changes, so no caller code is affected.

Test plan

• Manual: drop a WS handler task mid-stream ? after the default 30 s quiet window, is_turn_active(thread_id) returns False.
• Manual: a live stream that keeps producing chunks keeps its entry alive across the entire run.
• mark_turn_idle still removes entries immediately (the happy path is unchanged).