Report paths a skill writes that are not in the skill

[johnl-amd]

Summary

• Structural now reports a path a skill's markdown writes about its own files when that path is not in the skill. Two new pieces: references.path_mentions, which reads path-shaped tokens from raw markdown, and structure._path_errors, wired into skill_errors beside _required_errors.
• Why: a skill names its own scripts and data in prose and code spans, which are not markdown links, so the reference checks never see them. A path written against the source repo's layout resolves only for someone standing in that repo. It is dead in every way a skill is actually installed: vendored into a catalog, copied into an agent's skills directory, or fetched with the CLI. The agent follows it, finds nothing, and improvises, which is why a skill with one keeps passing its evals.
• Risk: low. No agent calls, no dataset or config change, nothing added to the schema. Deterministic, so it cannot flake.

What it does not flag

Every one of these started as a false positive against real skills and became a rule:

• Relative paths resolved against the file they are written in, the way a markdown link is, as well as against the skill root. Either reaches the file.
• ${SKILL_DIR}/scripts/launch.sh and similar, where the prefix is stripped and the tail is tested.
• Paths inside shell scripts, which resolve at runtime against the script rather than against the agent.
• Absolute paths, which are a container mount or a host layout.
• Files at the skill root, since a bare SKILL.md appears in install instructions for unrelated trees.
• evals/, which is test data and never the agent's to read.
• A path whose filename matches but whose tail does not, which is a different file that happens to share a name.

Reported only when the mention ends with exactly where the file ships, so what was meant is not in doubt.

Test plan

• 9 new tests in TestPathsASkillWritesAboutItself, one per rule above plus the failing case. Full suite 235 tests, green.
• Run against a real catalog of 10 skills: 2 findings, both genuine, 0 false positives.
• Verified in both directions. With the affected skill's paths left alone the check reports them; with the same paths made skill-relative, structural passes and the existing internal reference check still resolves 125 references across 43 markdown files.

Before merging

One published skill currently trips this. Its paths need fixing in the product repo it is federated from rather than in the vendored copy, which a re-import would overwrite. Since consumers pin the harness by tag, merging here does not reach them until a tag is cut and the ref is bumped, so this can land first, but a tag should wait for that fix.