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

Manual real-server proof; not a replay or unit test.
1. Install each commit with vp i; start T3CODE_DEV_SEED=0 vp run dev.
2. Run seed.py once in this directory. Add its project to each server:
   node apps/server/src/bin.ts project add --base-dir .t3 --title "Session lifecycle proof" <synthetic-project>
3. Choose an authenticated inexpensive Claude model via PROOF_MODEL.
4. Find each server PID by its listening port, confirming /proc/<pid>/cwd.
5. Run PROOF_MODEL=<chosen-model> python3 record.py before <base-checkout> <server-pid> <base-sha>
   Then the same command with after, head checkout/PID/SHA.

The driver starts one real parent turn, asks for three real delegate_task children,
observes read-only app projections and /proc, waits 15 seconds after parent completion,
and sends Child 1 the marker recall follow-up. There are no process mocks,
assertions, server patches, shortened timeouts, or thread settling commands.
The videos are sampled process-monitor output rendered at one frame per sample,
with actual elapsed seconds on each frame. Original sampling is approximately 1.1s.
Only app_owned delegations are relevant; native agents are excluded.

Private CLI responses are deliberately not included in the evidence bundle.
The public transcript includes selected run state and actual PID/RSS data only.
