[Docs] Add the 2026-09-23 rerun of the Google Flights benchmark

[opPO333]

Why

There was a concern that network latency to the TypeSafe and OpenRouter servers might delay the model's
responses. The first numbers were recorded in China, so we gathered a new set of data from a different
location to verify this.

How

We ran every arm again on 2026-09-23 from a Windows 11 PC. Google places this PC in Poland. Each arm ran
three times on the direct TypeSafe backend and three times through OpenRouter. The baseline is the runs
of 2026-09-18 and 2026-09-19 already in the doc. The arms were jev-ultrafast (A), the S1A policy on the
browser-use driver at tag jj-bu-baseline (B), and this repo on the stock Playwright MCP (C), the
no-settle copy (C') and the no-settle copy with batched actions (C''). All runs used the same Chrome
window, remote debugging on port 9222, and one scratch profile. Every arm searched for the same date,
October 25, 2026.

Before we edited the docs, a script checked all 27 result files. It checked the end status, the verified
flag, the final URL with the date in it, the number of decisions, and the median latency. We then ran
scripts/summarize_runs.py and scripts/compare_runs.py on the committed records to make the tables in
the doc.

What we found:

• The route to the model servers did not slow the model down. Each decision took 354 to 426 ms direct
  and 391 to 466 ms through OpenRouter. In the baseline it took 360 to 418 ms and 438 to 510 ms. The
  "waiting on the model" column is the same in both sets of runs. OpenRouter adds about 1 s per run over
  twelve decisions, as before. This PC is in Poland, so the test covers the route from Poland only.
• Page loading got much faster for the stock Playwright MCP arm. In the baseline the median run took
  42.7 s and only two of three runs passed. Now it takes 26.2 s direct and 28.9 s through OpenRouter. All
  three runs pass. Page waits dropped from 21.0 s to 11.3 s. Clicks dropped from 14.8 s to 9.1 s. The
  MCP server code is the same. Google's pages answered faster from here.
• Arm B took 8.0 s direct and 8.5 s through OpenRouter. In the baseline it took 11.8 s on 2026-09-19 and
  7.3 s on 2026-09-18.
• jev-ultrafast took 10.3 s, down from 14.1 s in the baseline. It threw away 5 to 8 decisions per run
  instead of 7 to 11.
• C'' is 1.3 s slower than its baseline. 0.7 s of that comes from the whole-page answer step (commit
  ec947e6). That step was added after the baseline runs.

Things a reviewer should know. The doc states all of them.

• Google showed a cookie consent page to this PC, as it does to any new browser in the EU. We clicked
  "Reject all" once before the runs. The timed window starts after navigation, so this does not touch
  any number.
• Google's menus do not open in a background tab on this Chrome build. jev-ultrafast opens its tab in the
  background, so its first three runs got stuck after one click. We changed one line so the tab opens in
  the foreground. Its decision loop, clock and model calls are unchanged. Every S1A arm already runs in
  the visible tab.
• Arm B's sidecar processes stay alive after a run ends. Two OpenRouter runs timed out until we killed
  the old processes. Those runs do not count. The doc lists every run we dropped.

Read docs/benchmarks.md first. The new parts are the sections "Rerun of 2026-09-23" and "Every arm in
one table".

What

• docs/benchmarks.md: a new section with the 27-row run table, the per-step waits of the median direct
  runs, and a short list of findings. The closing table now holds the baseline rows, the old
  jev-ultrafast row, and the nine rerun rows. It also has the command that rebuilds the table from the
  tree and the new chart. The heading "Six arms, one table" is now "Every arm in one table".
• docs/results/flights/rerun-2026-09-23/: the 24 S1A records plus compare.svg and compare-dark.svg.
  They sit in a subfolder so the old glob patterns still match only the baseline records. The
  jev-ultrafast records stay out of the tree, as before.
• README.md: the two links to the Flights comparison now mention the rerun. The README table compares
  agents against the chat model. The rerun does not change those numbers.
• CHANGELOG.md: one entry under Unreleased.

No code changed.

Verification

• uv run ruff format --check . && uv run ruff check . && uv run ty check pass.
• uv run pytest -q: 479 passed, 41 skipped, 3 failed. The three failures only happen on Windows and
  happen on main too. They are the exit code test for an unusable S1A_HOME and two socket
  timeouts in the BrowserHands tests. scripts/smoke.sh fails its first check on Windows because
  s1a list prints CRLF line endings. This branch changes only docs and data files.
• CHANGELOG.md and the docs say what the code does now.
• A script checked all 27 records before the docs were edited. The tables were rebuilt from the
  committed records with the repo's own scripts.

[Yunaik]

please don't push the results with 38k lines of code in here, but a compressed version or the code to reproduce the code.

[opPO333]

Done. I removed the raw JSON files and replaced them with a single compressed archive .tar.gz. The commit has been amended.

[Yunaik]

Thanks