# Busy thread lock reproduction fixture

This is a controlled-contention real-app proof. It reproduces the triage's explicit ThreadCommandExecutor.withLock(A) hold, using the real running server, fresh SQLite state, HTTP/MCP commands, ACP replay provider, WebSocket projections and web client.

The Python instrumenter adds only a filesystem-controlled lock holder to ThreadCommandExecutor's live layer, and externally controlled turn endings to the dev replay agent. The same instrumentation is applied to both commits. It does not alter Orchestrator.ts, the run-end reaction stream, queue promotion, or browser rendering.

Create two projects and two threads through t3.project and t3.thread. Start active replay turns on A and B, queue another message on B, open B and wait for its composer. Hold A's real thread lock, finish A, wait for its persisted terminal event, then finish B. Capture B while A's lock is still held. The assertion asks only whether B's queued follow-up produces its reply. The lock is released in finally, with a one-minute safety ceiling.

This fixture does not claim to reproduce a naturally slow send, delegated-child bookkeeping at production scale, or an actual provider's timing. It isolates the reported cross-thread queue blockage.

To reproduce, place the scenario in scripts/scenarios/ on current main, and invoke instrument.py APP_CHECKOUT from start-app.ts before starting each tested app. Pass a fresh control directory through T3_LOCK_PROBE_DIR and T3_SCENARIO_INHERIT_ENV=T3_LOCK_PROBE_DIR. Remove request, entered, release and released files before each side. Run scenarios:prove with base e367ef0da11e11a814911c93d812e2681c51294b and head 1235165531e7abe49286016786a87697789eac5e. No production data or paid model is used.
