research: name every flagged referenced file, not just divergent ones

[Ap6pack]

What the first scan showed

Run 34654103313 (60 skills, seed 20260816):

skills scanned:              58
  naming referenced files:   29   (50%, consistent with step 1's 47.2%)
referenced paths tried:      120
  fetched:                   107
  unreachable (named, 404):   13

verdicts on referenced files:
  CLEAN        104
  SUSPICIOUS     3

DIVERGENCE (SKILL.md clean, referenced file flagged): 0

An empty divergence set is a real result and it argues against the hypothesis that motivated this tooling: the unread surface is confirmed large, but in this draw nothing is hiding in it.

Why this change

Three files were flagged and the report named none of them. They fall outside the divergence set because their skill's SKILL.md was itself non-clean — documentation and code agreeing, which is the honest case, and the opposite of concealment. That is a reason to label them, not to drop them. As written, the only way to recover which skills, which paths, and which rules was to re-run the scan.

Two additions:

• All flagged referenced files, each with its parent doc verdict alongside the file verdict, so concordant and divergent cases are distinguishable on sight rather than by absence.
• Named-but-unreachable paths, grouped by skill. The 13 404s are already kept out of the clean count, but a reader could not tell which skills carry that unknown. "We could not read it" and "there was nothing to read" are different facts and only one of them supports a clean claim.

Timeout goes to 120 minutes so a 300-skill scan fits: roughly 900 registry fetches at the 0.55s pacing, plus rule-engine time on every fetched file.

Not changed

No detection logic, no thresholds, no extraction rules. Report-only, so the numbers above stay comparable across runs.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DNoTXU8k3pfSBzR7aJubqL

---

Generated by Claude Code