lint: stop --fix from destroying inline code and allowed-tools lists

[NagyVikt]

cue lint-skill --fix corrupts the files it edits. Measured over the 453-skill library, one run destroyed 723 inline code spans across 178 files and folded block-sequence allowed-tools into a single garbage string.

R009 — the em-dash fixer eats adjacent code spans

stripCodeAndFrontmatter is length-preserving: it blanks inline code to spaces so offsets stay aligned with the original. The fixer used that mask to find em dashes (correct) and to walk backwards over surrounding whitespace (not correct). A blanked `code span` reads as whitespace, so the walk crossed it and the slice dropped it:

-- `get_latest_news` — Fetch the most recent news articles
+-, Fetch the most recent news articles

The tool name is simply gone. Worst single file lost 44 spans.

Fix: the search still uses the mask; the whitespace walk now runs on the original content.

R005 — allowed-tools block lists collapse into one string

Two defects compounding:

1. It read the field with fmField, whose \s* crosses a newline. For a block sequence that returned the first bullet line, so - was parsed as a tool name → Bash(-:*).
2. Its fix rewrote only the key line, orphaning the remaining - item lines beneath a scalar. YAML folds those into the scalar.

research/trendradar went from a 17-entry list to one ~600-character string:

allowed-tools: Bash(-:*), Bash(mcp__trendradar__get_latest_news:*)
  - mcp__trendradar__get_latest_news
  - mcp__trendradar__get_trending_topics

That is still valid YAML — which is why a syntax-level integrity check waved it through. Only a semantic comparison (list of 17 → str) catches it.

Fix: R005 now reads both YAML shapes, replaces the whole construct, and leaves mcp__* ids and top-level tools bare rather than wrapping them. Bare Bash is a tool name, not a CLI name — wrapping it produced the nonsense Bash(Bash:*).

Why this shipped

The existing suite tested em dashes inside fenced code blocks but never inline spans. This PR adds 12 regression tests built from the shapes that actually broke, including the real-world lines above.

Verification

Check	Result
bun test src/lib/skill-linter.test.ts	91 pass / 0 fail
bun test (full)	3088 pass / 1 skip / 0 fail
--fix over 453 skills, inline spans lost	0 (was 723)
… allowed-tools list → scalar	0
… frontmatter unparseable / name: / description: / fences altered	0 / 0 / 0 / 0
lint result unchanged in substance	45 errors → 0, warnings 745 → 406

Also confirmed no new Bash(Bash:*) is introduced: 47 pre-existing occurrences, unchanged.

The corpus-wide application of the repaired fixer is opencue/skills#21.

🤖 Generated with Claude Code