fix(skill-creator): resolve Windows eval harness pipe reading and probe command leakage

[AbhiPra24]

Summary

• Replaced select.select() on pipes in run_eval.py with a thread-safe queue.Queue reader thread, avoiding WinSock-only select() crashes on Windows anonymous pipes.
• Updated find_project_root() to explicitly avoid matching Path.home() (~/.claude), ensuring probe command files are never written to the user's global command directory.
• Added cleanup_probe_commands() invoked at start and in finally blocks across run_eval.py and run_loop.py to guarantee probe files (*-skill-*.md) are safely deleted even when runs are interrupted or fail.
• Added explicit utf-8 encoding across file operations in run_eval.py, run_loop.py, improve_description.py, and utils.py to avoid Windows ANSI/cp1252 decoding errors.

Fixes #1692

cc @98zc5g5jyw-arch for review

[Hahaknight]

Verified on real Windows 11 (Python 3.13). This fixes defects 1, 2 and 4 from #1692 — but as written, defect 3 (twin collision capping recall at ~1/N) appears to remain, see last section.

Defect 1 (select on pipes) — fixed, verified. main still crashes in my environment: select.select([p.stdout], ...) raises OSError: [WinError 10093] (WSANOTINITIALISED — same class as the reporter's 10038: select() only supports sockets on Windows). This PR's thread-pump + queue.Queue pattern tested on the same platform: receives all subprocess output correctly. Cross-platform and keeps early-kill semantics.

Defect 2 (global ~/.claude/commands leak) — fixed, verified (3 cases):

• from a git-clone subdir with no .claude: returns the repo root (via the new .git marker), not $HOME ✅
• from a bare directory with no .git/.claude anywhere below home: returns cwd, never walks up into $HOME ✅
• cleanup_probe_commands(root, "myproj") removes only myproj-skill-*.md and leaves other-skill-xyz.md / unrelated.md untouched; the no-name variant sweeps all *-skill-*.md ✅ (good orphan cleanup for previously-interrupted runs)

Defect 4 (encoding) — fixed. run_loop.py eval-set read and all report/JSON writes now explicit utf-8, matching #1687/#1690.

Defect 3 (parallel twins → recall ~1/N) — not fixed by this PR. run_single_query still writes {skill_name}-skill-{uuid}.md into the shared project_root/.claude/commands/, so while ProcessPoolExecutor runs ~10 probes concurrently, every claude -p session sees ~10 near-identical twins and invokes an arbitrary one — only the probe owning that UUID counts a hit. That's exactly the 11% train / 8% test plateau (~1/10) the reporter measured, which persisted "regardless of description". The reporter's local fix (per-probe tempfile.mkdtemp() root + rmtree in finally) took the same description from ~11% to 60/60 — worth adopting here or in a follow-up; it would also make the defect-2 leak structurally impossible rather than mitigated.

Minor: the reporter's last suggestion (loud failure when a large share of probes error, instead of feeding recall=0% into the improver) isn't addressed — a one-line guard in run_eval raising if e.g. >50% of futures raised would stop quota burn on invalid measurements.