skill-creator: description-optimization (run_loop) is broken on Windows - claude not found + select() on pipes

## Summary
The skill-creator's description-optimization tooling (`scripts/run_loop.py` →
`scripts/run_eval.py` + `scripts/improve_description.py`) cannot run on Windows.
Every `claude -p` subprocess fails with `FileNotFoundError: [WinError 2]`, and
the streaming reader uses `select.select()` on a subprocess pipe, which is
POSIX-only. On Windows all queries error out and the loop crashes when it
reaches the "Improving description" step.

## Environment
- Windows 11, Python 3.14
- Claude Code CLI 2.1.158 (installed via npm; resolves to `claude.cmd`)
- skills repo at commit `da20c92`

## Reproduction
```
python -m scripts.run_loop \
--eval-set \
--skill-path \
--model --max-iterations 1 --verbose
```
Result: `Warning: query failed: [WinError 2] Das System kann die angegebene Datei nicht finden`
repeated for every query, then a traceback ending in
`FileNotFoundError: [WinError 2]` from `improve_description.py`.

## Root causes & fixes

**1. `subprocess` can't launch the CLI by bare name `"claude"` on Windows.**
On Windows the CLI is `claude.cmd`; `CreateProcess` cannot execute a `.cmd`
directly and bare `"claude"` isn't resolved. Fix: resolve via
`shutil.which("claude")` and, on `os.name == "nt"`, route through the command
interpreter (`%COMSPEC% /c ...`). Applies to both `run_eval.py` (line ~71) and
`improve_description.py` (line ~26).

**2. `select.select()` on a pipe is POSIX-only (`run_eval.py` line ~108).**
On Windows `select` only works with sockets, so the streaming read loop can
never read the subprocess stdout. Fix: drop `select`; iterate
`for raw_line in process.stdout:` (portable) and enforce the timeout with a
`threading.Timer` watchdog that kills the process. Early-trigger detection logic
is unchanged.

## Patch
A unified diff against `da20c92` is attached (`skill-creator-windows-fixes.patch`),
touching only `scripts/run_eval.py` and `scripts/improve_description.py`. After
applying, `run_loop` launches `claude -p` successfully and streams events on
Windows.

## Note (separate, lower priority)
With the subprocess fixed, the trigger-detection harness still reported
`recall=0%` on Windows: the temp slash-command written to
`/.claude/commands/` did not surface as an invocable Skill under
`claude -p` in this environment, so the model never emitted a `Skill`/`Read`
tool call for it. Worth confirming whether project-scoped commands are exposed
to `claude -p` headless runs on Windows, as the description-optimization signal
depends on it. (Filed separately from the two subprocess bugs above, which are
clear-cut.)

[skill-creator-windows-fixes.patch](https://github.com/user-attachments/files/28425743/skill-creator-windows-fixes.patch)