fix: per-file fix reporting, BOM-safe contents insertion, stale-offset guard

[MaximeGaudin]

Summary

A cluster of correctness/atomicity fixes around --fix:

• Per-file fix reporting (--fix: a write failure partway through a multi-file run leaves earlier files already mutated on disk, with no report printed and no indication of what was changed #227): a write failure partway through a multi-file run used to abort before the report was rendered, leaving earlier files fixed on disk with only a bare error line behind. fix::apply now records a per-file failure (Applied::failed) and continues with the remaining files; the CLI turns each failure into a note on the report and exits with the failed code (3), so the reader sees exactly which files changed and which did not.
• BOM-safe contents insertion (--fix: a UTF-8 BOM at the top of a file breaks bundle/contents-list's title detection, so the generated Contents block is inserted before the document title instead of after it #230): with_contents() read the file's first line with a UTF-8 BOM attached, so the title never matched and the Contents block was spliced in above the mark and the title it belongs under. The first line is now scanned past the mark — the same tolerance skill.rs applies to frontmatter — and the mark's bytes stay on disk untouched.
• Stale-offset guard (--fix: no hash/mtime guard between when a Fix's byte offsets are computed and when they are applied — a library-level TOCTOU that the CLI happens to avoid only because lint-then-fix runs in one process #234, --fix applied against stale byte offsets (file changed between the lint pass and the fix pass) is untested #247): fixes are byte offsets into the text they were measured against, but apply_to_file() re-read the file and trusted the offsets blindly — a file rewritten between the lint pass and the fix pass (even to the same length) had the fix spliced into bytes nobody measured. The report now records a fingerprint (length + cheap hash) of each fixable file's text when it is read, and apply() re-checks before writing: a mismatch skips that file with a note instead of corrupting it.
• Read-only file coverage (No test for --fix writing to a read-only file #277): pins both halves of --fix behaviour on a read-only SKILL.md — it is still fixable by replacement (rename asks the directory for room, and keeps the mode), and when the directory refuses too, the failure is reported and other files still get fixed.

Fixes #227
Fixes #230
Fixes #234
Fixes #247
Fixes #277

Test plan

• test: reproduce #227 — e2e: multi-file --fix with one unwritable directory failed before the fix (no report on stdout); passes after
• test: reproduce #230 — BOM'd long reference put ## Contents above the title; passes after
• test: reproduce #234/#247 — same-length rewrite between compute and apply corrupted the file (fix applied to wrong bytes); skipped with a note after
• test: cover #277 — verified red at c226adf (before the --fix: a write failure partway through a multi-file run leaves earlier files already mutated on disk, with no report printed and no indication of what was changed #227 fix), green at HEAD
• Local checks mirror CI: fmt, clippy -D warnings, cargo test --workspace, cargo deny check, no-todos, rules-json, MSRV cargo check --locked
• CI green on the PR

Note: a_later_path_with_its_own_config_is_named_when_the_first_path_governs is a pre-existing flake reproducible on origin/main (XDG_CONFIG_HOME env mutation racing parallel tests); it is unrelated to this diff.