PR #32261 real runtime verification
baseline: github/main f31ce3eeb5f4adbfe7f961403822cf90db05808f
PR head: 30ff7d83d4c885b6699d02b7b258e997d2e40388

Main: vp i; background vp run dev, server 16733 / web 8693.
Playwright 1.64.1 + /usr/bin/google-chrome, real Codex gpt-6.1-sol.
Created a project in the client pointing to /home/user/token-usage-verification.
Created a thread and sent two prompts through its composer:
1. Read README.md using a shell command, reply Token usage verification ready.
2. Run git status --short and report whether the checkout is clean.
Both completed. Main's surface picker has no Insights entry.
Stopped main by captured process group before starting the PR.
Copied the stopped app-created development state to preserve matching data.

PR: vp i; background vp run dev --home-dir /home/user/t3code/.t3,
server 13773 / web 5733. Opened the same thread and clicked Insights.
Actual client server.priceUsage request and successful response:
turn 1: uncachedInput=7330 cacheRead=25728 cacheWrite=0 output=38
         total=33096, costUsd=0.0176128, costSource=modelPriced
turn 2: uncachedInput=7593 cacheRead=25728 cacheWrite=0 output=53
         total=33374, costUsd=0.0182888, costSource=modelPriced
UI: 66.5K tokens, 2 turns, $0.04 estimated API cost.
Click By turn > Turn 2: aria-pressed=true on Turn 2 flame-chart bar;
detail shows 33.4K tokens, $0.02, model gpt-6.1-sol.
Focus, zoom in, fit, clear, and select Turn 1 all worked (recording).

Fresh PR provider turn: read README.md, reply PR head provider path verified.
Completed through the real Codex adapter in 7.7s.
turn 3: uncachedInput=1926 cacheRead=34560 cacheWrite=0 output=39
         total=36525, costUsd=0.007698, costSource=modelPriced
UI updated to 3 turns, 103K tokens, $0.04 estimated API cost.
Exact thread totals: 102995 tokens, costUsd=0.0435996.

Sent server.updateSettings over the authenticated development WebSocket,
aliasing gpt-6.1-sol to a model absent from the rate table.
Same real turns: all three costs displayed Unknown; header omitted cost.
Added override on the alias target: input=1, output=2, cacheRead=0.1,
cacheWrite=1 dollars per million tokens. Actual server.priceUsage results:
turn 1: 0.0099788; turn 2: 0.0102718; turn 3: 0.0054600.
Each matched independent arithmetic from the actual client token payload.
UI refreshed to $0.03. Restored the original empty aliases/overrides;
UI returned to $0.04 without altering provider usage records.

Nearby behavior:
- Existing Usage Tokens view loads actual Codex session totals and costs.
- Files surface opens the real README.md source alongside Insights.
- Close/reopen Insights and reload preserve correct three-turn data.
- A second composer-created thread ran a real Codex turn and showed only
  its own 1 turn / 16.5K tokens / $0.02. Returning to the first thread
  still showed 3 turns / 103K / $0.04.
- Inspected light 1440x1000 and dark 1280x900 renders.
- No browser page errors observed.
- Uploaded 6.52s interaction video played to completion and sought to 3s
  in T3's preview player: readyState=4, 1440x1000, no media error.

No code changes, commits, or pushes were needed.