test(skills): cover Windows catalog

[orangeCatDeveloper]

Fixes #3901

Summary

Windows could report a healthy Skill Catalog suite while six filesystem, recovery, and multi-client contracts never ran.

Five tests returned before their assertions on Windows, while the real two-client transport test carried an explicit Windows skip. The tests now use Windows ACLs for unreadable fixtures, accept the Windows process-termination representation, exercise the existing named-pipe transport with an isolated skill home, and retain the open-handle and symlink recovery assertions.

The blocking Windows recovery lane runs every Skill Catalog test file and requires exactly 90 passes with zero skips; its CI contract prevents the file list or count from silently shrinking.

Verification

Before:

447:  if (process.platform === 'win32') return;
466:  if (process.platform === 'win32') return;
216:  if (process.platform === 'win32') return;
540:  if (process.platform === 'win32') return;
625:  if (process.platform === 'win32') return;
44:  skip: process.platform === 'win32',

After (Windows recovery run):

# tests 90
# suites 1
# pass 90
# fail 0
# cancelled 0
# skipped 0
# todo 0

AI use

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Codex implemented the cross-platform test fixtures and Windows CI gate, then performed local validation and an independent subagent review.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[Astro-Han]

Approve — the Windows adaptations are real execution, not a different way of skipping. I verified each: icacls /deny genuinely denies read (the rejects assertions are unchanged), the SIGKILL case only swaps the process-termination representation while keeping the transaction-directory and recovery assertions intact, the open-fd case is honestly renamed to "stale save" and still asserts commit_outcome_unknown plus the frozen content, and the two-client test now exercises the named-pipe transport with an isolated skill home.

Two suggestions, neither blocking:

• The fixture change at skill-catalog-repository.test.ts:323 trades coverage rather than gaining it, and the body does not mention it. The id went from a newline (bad\nskill) to DEL (bad\u007Fskill) because Windows forbids 0x00-0x1F in filenames while DEL is allowed. But this test is named "noncanonical external ids remain wire-safe governance entries", and a newline is precisely the character wire-safety cares about - it can split or inject lines in a line-oriented representation, which DEL cannot. POSIX previously covered the dangerous character and now covers a harmless one. Keeping both would be strictly better: the newline case gated to non-Windows, plus the DEL case everywhere.
• ci-test-plan.test.mjs asserts # tests 90 / # pass 90, duplicating a number the workflow already hardcodes. The real guard is the workflow's own PowerShell check, so the meta-test copy adds no independent protection while making a legitimate new test require two edits. The adjacent "root initialization replacement race" test asserts only # skipped 0, which is the better precedent.

AI-assisted review disclosure: Claude Code compared the before/after skip sites and the CI contract; I verified the ACL mechanism, the retained assertions, and the fixture-character trade against current main myself.