audit-dead-code: detect unreferenced files in any language

[kyle-sexton]

Summary

code-tidying:audit-dead-code 0.23.2 advertises "orphaned files" (SKILL.md:2). SKILL.md:36 limits
that to TS/JS, and in practice it also needs a package.json root because it comes from knip. In a
scripts-only repo, an unreferenced script is the likeliest form of dead code, and no lane checks for
it. This covers shell, PowerShell, Python entry points, standalone JS, and files in languages with
no lane.

In melodic-software/.github, a manual check proved each script file alive by a basename reference
search. The references came from .claude/settings.json:20, .cursor/environment.json:4, and
.github/workflows/ci.yml:127. The skill has no equivalent.

Fix

• Add a language-agnostic unreferenced-file shape at T2. It reuses the grep lane's repo-wide
  reference search, keyed on the file basename and its repo-relative path, and is limited to
  source-language files.
• Count a reference from CI workflows, settings, manifests, and docs as alive evidence. That
  evidence is what saves scripts that are invoked only by path.
• Name the known blind spot: a file loaded by a computed path or glob. The report should list it as
  uncertain, not dead.
• Update SKILL.md:2 and SKILL.md:36 to state the new coverage.

Verification

• Done when a fixture with one script referenced from a workflow and one referenced nowhere yields
  exactly one unreferenced-file candidate.

Related

• Found in a playbooks:repo-sweep sweep of melodic-software/.github (PR chore(skill-quality): harden static checker + record terminal retrofit scope #153).
• Sibling issues from the same audit: coverage accounting, JS outside a package.json root.

[kyle-sexton]

This comes from the 2026-09-29 audit of the Cursor agent run.

The delivered orphaned-file detection covers only extensions the classifier lists, not "any language": dc_lang_of_path returns other for .s, .sol and extensionless scripts, so an unreferenced one is never reported. The owner already acknowledged this as out of scope in the #5085 review reply. The gap is now tracked in #5329 (extensible classifier, with acceptance criteria).

PR #5327 (draft, not yet merged) also makes these changes here: it rewords "any source language" to "source files with a recognised extension" in the skill description and SKILL.md, skips a path containing a tab or newline in the unreferenced-file scan (this fixes the tab-in-path owners.tsv row), and corrects the fixture comment. This issue stays closed.