perf(test): measure the Windows column of the wall-clock baseline and the gate, affected-run and corpus targets #10241

Parent: #3716 (slice 2 of 3). Blocked by #10200 (slice 1).
Runs in: HOST. Needs a healthy Windows 11 Git Bash host; the lane runs on Linux; the host runner works it.

Problem

#3716 criteria 1 to 4 are Windows measurements: the Windows column of the baseline table, no gate script over 60 s, scripts/affected-tests.sh --run on a typical 50-file diff under 2 minutes, and the full scripts/run-plugin-tests.sh corpus under 10 minutes with the number recorded. No current Windows number exists; the last ones are from 2026-09-04 to 2026-09-08 (#3716 comment), before many optimizations.

Fix

On a Windows host that does not carry the #3715 process-creation condition (or record that it does), at a stated origin/main commit:

1. Run the full corpus with slice 1's timing option and fill the Windows column of docs/test-wall-clock-baseline.md, with the total.
2. Time each gate script slice 1 lists in its typical single-diff invocation.
3. Run scripts/affected-tests.sh --run over a real merged 50-file diff (name the commit range) and record the time.
4. Open one PR with the doc update. In the PR body list every gate script over 60 s, and say whether criteria 3 and 4 are met.

Done when

• docs/test-wall-clock-baseline.md has a filled Windows column, the host's OS build, Git Bash version and the commit measured
• The doc records the full-corpus Windows total and the 50-file affected-tests.sh --run time with its commit range
• The PR body lists each gate script over 60 s on Windows, or states there is none
• A comment on test-perf: diagnose and cut the wall clock of the local test and gate corpus, especially on Windows Git Bash #3716 states which of criteria 2, 3 and 4 are met, so per-offender slices can be filed for the rest
• ci-status is green on the PR head

Split by the backlog HITL session (2026-10-25, owner rule Q93: split large work before a run picks it up). This was generated by AI.