Proof of unchanged npm Codex restart caching on Windows Server 2025.

Prepare two T3 worktrees at the SHAs recorded in measurement.txt; run vp i in each.
Install @openai/codex@0.171.0 with npm, and put its codex.CMD directory on PATH.
Keep measure.cjs and spawnlog.cjs together. Use fresh synthetic state, never live state.
Warm both sides with alternating real server launches. The recorded exploratory batch
contains nine launches per side. The logger was finalized during this warm-up batch.
Run the final fixed seven-pair batch using those warmed state directories:
  node measure.cjs <base-checkout> <head-checkout> <new-output-dir> <warmup-output-dir>
Capture stdout as final.jsonl. Each launch uses node --require spawnlog.cjs src/bin.ts
serve --base-dir <side-state> --port 4800 --host 127.0.0.1 from apps/server.
T3CODE_DEV_SEED=0 is set. After 25 seconds taskkill /PID <captured pid> /T /F stops it.
  python analyze.py final.jsonl

measurement.txt is the complete measured output. final.jsonl retains the raw measured
events, raw.jsonl the exploratory warm-up events. No measured runs were excluded.
The worker interval begins at preload boot of the same worker that spawns app-server,
and ends at that spawn. Whole-server module loading is excluded. Verification duration
sums the two completed child-process checks. No verification launches means zero cost.
All 14 measured app-server probes initialized Codex 0.171.0, returned an unauthenticated
account requiring OpenAI auth, and exited with code 0. No paid agent turn was run.

The proof covers unchanged npm Codex restarts only. It does not exercise fingerprint
invalidation on binary/package updates, non-npm launchers, or desktop shell behavior.
