Queue ordering runtime evidence

Both sides run the real Linux development server from separate Git worktrees, with fresh databases and identical synthetic project/thread data. No unit test or mocked orchestrator is used. The CLI authenticates against each server and calls its real HTTP MCP endpoint. Projection observations use read-only SQLite connections.

Base: d290c1f1b6c93416dfda6c9f7cd15d928a9aeb4c
Head: a84725a74daf0a5d8c8a6ccf6ebeb8b012670a6d

Reproduce separately for each commit:
1. Install dependencies with vp i.
2. Run python3 instrument.py <checkout> once before starting the server.
3. From the checkout, run T3CODE_DEV_SEED=0 vp run dev:server against a fresh worktree-local .t3.
4. From the checkout, run node apps/server/src/bin.ts project add --base-dir .t3 . --title 'Queue order evidence'.
5. Run python3 driver.py <checkout> before (or after).
6. Run python3 render-recording.py <timestamped-output-directory> to render the timestamped terminal recording as an MP4.

Identical temporary instrumentation on both sides:
- A 20-second delay at the start of the terminal run-update handler widens the reported run-completion / queue-release race window. It does not alter dispatch decisions or queue release logic.
- The free replay agent waits three seconds on each prompt, so QUEUED can be sent while ACTIVE is still running.
- The CLI send command uses MCP mode=auto for the exact synthetic message NEWER and mode=queue otherwise. The original CLI exposes only queue mode; this thin change exposes the actual MCP auto path without changing the server.

The driver starts ACTIVE, sends QUEUED while it is active, observes ACTIVE completed with QUEUED still queued, then sends NEWER in auto mode. It records delivery responses and real projected status changes until all turns finish, then prints execution order from persisted startedAt values. Queue promotion can change ordinal; labels follow userMessageId.

MP4 files render the captured terminal events at their original timing, with a six-second final hold, identical 1200x720 dark canvases, and no edited observations. The original timestamped recordings are also supplied as .cast files; plain text transcripts make the output searchable. No UI screenshot is claimed.

Observed base: ACTIVE -> NEWER -> QUEUED.
Observed head: ACTIVE -> QUEUED -> NEWER.

Scope: proves the eligible-queue race with controlled handler delay, not its frequency under natural load. Failed-worktree holds and other dispatch modes were not exercised.

AI-assisted; verified with the tests and tools listed above.
