fix(skill-creator): run_eval trigger detection misses real skill name and bails on first non-Skill tool

[Polluelo978]

Problem

scripts/run_eval.py::run_single_query fails to detect that a skill triggered, causing the description-optimization loop (run_loop.py) to report recall=0% for every should-trigger query. With all candidates tied at 0, the loop just returns the original description and never actually optimizes.

Two root causes in the stream detection:

1. Only the registered command name is matched. The function registers a temp command named {skill}-skill-{uuid} and only checks whether that random name (clean_name) appears in the tool input. But an installed skill fires under its real name (e.g. cve-intel), so the random name never matches -> always False.
2. Bails to False on the first non-Skill/Read tool. Real Claude Code runs often call Bash/WebFetch/etc. before consulting the skill. The content_block_start handler does return False on the first such tool, missing a later Skill/Read call.

Fix

• Match the real skill_name in addition to the registered clean_name, in both the streaming path and the full-assistant-message fallback.
• Do not return False on the first non-Skill/Read tool. Reset the pending state and keep scanning until type == "result".

Early-return on a positive match keeps positives fast; negatives run to result (short Q&A), bounded by --timeout.

Evidence (same skill + eval set, 3 runs/query, --model claude-sonnet-4-6 --timeout 120)

metric	before	after (held-out test)
recall	0%	100%
precision	100%	100%
accuracy	50%	100%

After the fix the loop can measure description quality as intended (held-out test 8/8).

Notes

• No behavioral change for skills not installed locally: clean_name matching is preserved; skill_name matching is additive.
• Single-file change in skills/skill-creator/scripts/run_eval.py.