Usage-limit cleanup proof

This is a manual server/RPC reproduction, not a unit test or a web UI capture.
Use isolated dependency-installed checkouts at the base and head commits.

Place driver.ts, record.py, and render-recording.py in one empty output directory.
Run:
  python3 record.py <base-checkout> <output-directory> before 14732
  python3 record.py <head-checkout> <output-directory> after 14731
  python3 render-recording.py <output-directory>

Each invocation creates a new timestamped synthetic T3 home and a real git repository
with a clean branch-backed linked worktree. It seeds production ProjectionStore and
ProjectStore projections offline through their apply methods, then starts the real
T3 server with T3CODE_DEV_SEED=0. No test fixture or mocked service is used.

State: failed six days ago, usage-limit reset five days ago, no armed recovery.
Automatic usage-limit resume and snooze are enabled. Automatic settlement is disabled.
No provider process is invoked; the synthetic provider/model selection is named demo.
The exported limitRecoveryCommand returns null from the real recovery candidate.

The driver authenticates locally, reads the running server's shell over HTTP, subscribes
to storage cleanup reports over its production WebSocket RPC, enables the one-day
inactive-worktree policy via server.updateSettings, and calls server.runStorageCleanup.
The directory's real filesystem existence is reported alongside each outcome.

Base: automatic and manual cleanup keep the directory as pending work. As an additional
control only while that directory exists, disabling auto-resume then running cleanup
removes it as inactive for six days.
Head: the settings-triggered automatic sweep removes it as inactive for six days before
the explicit cleanup RPC, so that later RPC has no remaining entry for this directory.

The text files contain captured command output. The cast files preserve output arrival
timestamps. PNGs render that output; MP4s replay those same timestamped transcript events
at 1280x900 with the same dark terminal presentation. They are not browser recordings.
The source driver and renderer are included so this presentation is reproducible.
All data are synthetic. Startup logs and local auth tokens are deliberately not published.
Servers stop in finally blocks; remove the temporary evidence-driver.ts from each checkout.

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