Usage pricing race proof

These are real WebSocket RPC responses from separate unmodified server checkouts.
The PNGs and MP4s render the captured RPC transcript; they are not browser or terminal screen recordings. MP4 event arrivals use the timestamps in the JSON and end with a reading hold.

base: 9e19b7f7e1ab504e05ee73df51c7fc7700f246af
head: 24cf23b298d0dc8e15e19e8333e9088a144b1bee
Linux, Node 26.12.1. Fresh databases, isolated HOME, same synthetic Codex file and identical saved rate table. No real provider or account used.

Reproduction driver:
1. Put seed.py and drive.mjs in a new evidence directory. Run python3 seed.py.
2. In each checkout, install dependencies with vp i.
3. Set HOME to a new empty directory (important: prevents scanning unrelated histories), CODEX_HOME to the generated codex directory and T3CODE_DEV_SEED=0.
4. Start node apps/server/src/bin.ts serve --base-dir <evidence>/<side> --host 127.0.0.1 --port <port>, saving startup output to <evidence>/<side>.log. Default driver ports: base 18271, head 18272. Keep the pairing logs private.
5. Run node drive.mjs <side> <port>. Each side gets its own fresh state directory and initial saved rates; never rerun against refreshed state for a before/after proof.

The captured runs used side names base-clean and head-clean with ports 18273 and 18274. They shared the generated synthetic file and the same saved table. Initial and control costs cover all 400000 records; no other usage buckets existed.
A reads the last 30 days, then starts a second non-provisional read with its window shifted one day. B starts refreshUsageRates 300 ms later. A final read shifts another day; all windows include the synthetic records.

Observed base: B finished at 09:41:45.101Z, A at 09:41:45.261Z. A's $1635 cost was labelled fresh with 09:41:45.011Z, but the control using that fresh table cost $163.50.
Observed head: B finished at 09:41:51.097Z, A at 09:41:51.221Z. A's $1635 cost retained cached/07:40:00.877Z. Control: $163.50 fresh/09:41:51.016Z.

The saved-first UI path was not exercised. This proves the server metadata race only.