skill-creator: description optimizer reports scores it never measured (recall structurally 0%) + 2 Windows blockers

```json
{"name":"Skill","input":{"skill":"runtime-truth","args":"..."}}
```

So the skill and description were fine all along — the harness could not see it. This is the most serious issue: **the tool silently reports a metric it cannot compute**, and a user following the documented workflow would apply an unvalidated description believing it was measured. I could not verify whether macOS/Linux behaves differently, but the registration/detection mismatch looks platform-independent.

**Suggested fix:** register the candidate as an actual skill (e.g. a temp skill dir with `SKILL.md` frontmatter carrying the candidate description) so it appears in `available_skills`; or, if the slash-command approach is intentional, also count `SlashCommand` invocations. Either way, a guard that errors out when recall is 0% across all candidates would have surfaced this immediately instead of emitting confident numbers.

---

## 2. `select.select()` on a pipe — Windows blocker

```python
# run_eval.py ~L108
ready, _, _ = select.select([process.stdout], [], [], 1.0)
```

On Windows `select.select()` accepts **only sockets**, so every query raises:

```
Warning: query failed: [WinError 10038] An operation was attempted on something that is not a socket
```

Every query fails, and because failures are swallowed as warnings the run still produces a scored report.

**Suggested fix:** read the pipe on a background thread into a `queue.Queue` (preserves incremental streaming and early trigger-detection short-circuit), or use `process.communicate(timeout=…)` if streaming isn't required. Patch that worked for me:

```python
import threading, queue as _queue
_chunks = _queue.Queue()

def _pipe_reader(fd, q):
try:
while True:
data = os.read(fd, 8192)
if not data:
break
q.put(data)
finally:
q.put(None)

threading.Thread(target=_pipe_reader,
args=(process.stdout.fileno(), _chunks), daemon=True).start()

while time.time() - start_time < timeout:
try:
chunk = _chunks.get(timeout=1.0)
except _queue.Empty:
if process.poll() is not None:
break
continue
if chunk is None:
break
buffer += chunk.decode("utf-8", errors="replace")
```

---

## 3. `UnicodeEncodeError` writing the HTML report — Windows blocker

```
File "scripts/run_loop.py", line 151, in run_loop
live_report_path.write_text(generate_html(partial_output, ...))
UnicodeEncodeError: 'charmap' codec can't encode character '✗' in position 14035
```

`Path.write_text()` uses the locale encoding (cp1252 on Windows), and the report contains `✗`/`✓`.

**Suggested fix:** `write_text(..., encoding="utf-8")` at every report/JSON write site (`--report none` does not avoid it, since the *live* report is written regardless).

---

## 4. Minor: orphaned command files are left behind

`run_eval.py` unlinks the temp command file in a `finally`, but when the run dies (crash in #3, or the user interrupts) the files survive. After my runs, 10 files remained in `.claude/commands/`:

```
runtime-truth-skill-199fa8fd.md
runtime-truth-skill-27daf233.md
… (10 total)
```

They then show up in `available_skills` in *every subsequent session*, cluttering the list and competing with the real skill during triggering. Worth a startup sweep of stale `*-skill-.md` files, or writing them to a temp dir outside `.claude/commands/`.

---

## Impact

Issues #2 and #3 make the description optimizer unusable on Windows out of the box. Issue #1 is worse than unusable: it produces confident, well-formatted metrics for something it never measured, and the documented workflow ends with "apply `best_description`" — so the failure mode is silently shipping an unvalidated description.