test(windows): guard clean MCP stdio startup - #2952

What does this PR do?

Adds a native Windows regression guard for #654: a successful MCP initialize + tools/list exchange, run against an isolated cache and runtime directory, must leave nothing on stderr except the server's own structured log lines.

• tests/windows/test_mcp_stdio.py (new): the guard.
  • The English and Spanish "path not found" messages reported in [Windows] MCP stdio startup emits localized path-not-found message on stderr #654 are named cases, so a red run says exactly which known leak came back.
  • Any other non-empty stderr line that is not a structured level=<name> msg=<event> record fails as well, so the same leak on a Windows machine in any other language is caught without listing its translation.
  • The log format is pinned to text (CBM_LOG_FORMAT=text), so an inherited JSON setting cannot change the verdict.
  • Cache and runtime live under one temp root, so both are removed together, also when the run fails, after the daemon is stopped.
• tests/windows/mcp_stdio.py: close() joins the stderr reader, and the new stderr_complete() reports whether the reader reached EOF. The guard reads stderr only after it has. A reader still running after the server exited is a SETUP result (exit 2), never a verdict on a partial capture; the wait there is a liveness backstop only and never decides pass or fail.
• scripts/test-windows.ps1: registers the guard in the green guard list and documents it in the header.

Evidence

• Hosted CI: the Windows guards job runs this guard on this PR (the change selects that lane).
• Pending before merge: a run on our local Windows 11 ARM64 machine with a native build of this branch on current main, covering the guard list, a .cmd wrapper that prints the Spanish message (expected red), a German-message wrapper (expected red only with the raw-line rule), a wrapper that keeps stderr held open (expected SETUP), and a cleanup check with an exception injected after initialize (no cbm-mcp-stdio-* directory left). The results will be posted here.
• Locally on macOS: the stderr classifier puts the Spanish and English lines into their named cases and a German line, a JSON line and a bare level=info into the raw-line case, and a clean log yields nothing; stderr_complete() is false while a background child holds stderr open and true once it is released (driven through a FIFO, so the result does not depend on timing).

Credit

Distilled from #2217 by @lorenzozanee, with the cleanup by @Shinken009 (both Co-authored-by on the commit); the original could not merge because its commit has no DCO sign-off, which only the author can add. #2213 was an earlier duplicate of the same check.

Refs #654

Checklist

• Every commit is signed off (git commit -s)
• Tests pass locally (Windows run pending, see Evidence)
• Lint passes (no C sources changed; the PowerShell file stays pure ASCII)
• New behavior is covered by a test (the guard itself)