fix(drift): resolve MISSING_PATH false positives from #143

[theyashasvipandey]

What

Fixes the three MISSING_PATH defects reported in #143, each of which still
reproduced on 0.8.x:

1. Negation was heading-scoped only. It is now also scoped to the sentence
   containing a reference. The sentence — not the line — is the unit that
   governs: markdown wraps prose freely, so a cue and the reference it applies
   to routinely sit on different lines of one bullet.
2. A bare trailing directory was excluded from the unrooted-reference guard
   by its own separator. The separator now marks a directory reference
   instead of disqualifying the value, so screenshots/ is prose when nothing
   by that name exists, while .mex/local/ still roots at a real directory and
   remains a claim the checker tests.
3. The fallback filename globs lacked dot: true, so a file documented by
   its bare name was never found inside a hidden directory such as
   .github/workflows/.

The numeric-delta, overall/overall and glob cases from the same report
already pass through the isUnrootedReference guard added since 0.7.x and
needed no change. All are covered by tests so they stay fixed.

Deliberate divergence from the report

The reporter also proposed demoting surviving MISSING_PATH issues from error to
warning. That is not included, per the maintainer reasoning in the issue
thread: demotion does not remove a false positive, it only makes it cheaper, so
it takes ~34 issues to floor the score instead of ~10 — bigger bucket, same
leak. It also answers the open question left for the reporter: all three
defects turned out decidable, so each one drops the claim before it becomes an
issue and nothing needed demoting.

An approach that was tried and rejected

Scoping negation to the whole paragraph passes all the same tests, but an
audit across this repository's scaffold showed it suppressed 7 of 167 path
claims where only 1 was a true negation — including four real, existing
directories, silenced because an unrelated clause elsewhere in the same block
mentioned a removal. A false negative is worse than the false positive this is
meant to kill. Sentence scope brings that to 2 of 167.

Known residual

One false negative remains: a sentence that both names a path and contains a
cue after it (… implemented on \X`: successful agent exit no longer
authorizes …) is still suppressed. Making a colon a sentence boundary would fix it and break the reporter's primary case (Deleted orphaned files:
`check_commands.js``), so it is left as the accepted cost of keyword-based
negation.

Why

Closes #143.

mex check reported errors for files that were intentionally absent or were
never paths at all, which zeroes drift scores on healthy repositories.

Two things this PR does not claim:

• It does not move real-world scores. Verified against five local scaffolds
  with real .mex/ directories: 0 issues removed, 0 introduced on every
  one. A clean 100/100 scaffold stays 100/100 — the safety property the
  reporter observed across their known-good projects — but none of these
  projects contain the defects MISSING_PATH false positives: negation only heading-scoped, non-path tokens treated as paths, dot-dir glob blind spot #143 describes. The false positives dominating
  them are different classes, which will be filed separately.
• It does not change severity, scoring, or any on-disk format.

Type of change

• Bug fix
• New feature
• Refactor
• Docs
• CI/Tooling

How to test

1. npx vitest run src/drift — 10 pass, 1 skipped. These are the first tests
   under src/drift, which had no coverage across its 19 source files.
2. The skipped case documents an out-of-scope defect: a version-shaped
   reference such as release/2.1.0 or python/3.11 reads as a file extension
   to the same guard and is still reported. Tracked separately; unskip with
   that fix.
3. On a scaffold, check the before/after issue list rather than the score:

   node dist/cli.js check --json
   

   On this repository the raw deduction moves from −163 to −150 (two issues
   removed, none introduced). The score reads 0/100 either way, because it is
   floored and dominated by staleness.
4. Regression guard worth reading: still checks a directory reference rooted at a directory that exists pins that the trailing-separator change did not
   quietly turn every documented directory into prose.

Checklist

• Tests pass (npm test) — not run in full locally. tsc --noEmit is
  clean and the targeted suites (src/drift, src/wiki/migration,
  src/__tests__, src/agent-skills) pass. The two agent-skills/installer
  failures seen locally are pre-existing and reproduce identically on stock
  main. Leaving this to CI.
• No breaking changes — no public API, severity or format change.
  isNegatedText is a new internal export of src/markdown.ts and is not
  re-exported from src/index.ts.
• Tested locally with a real project — five scaffolds, stock vs. patched
  builds compared issue-by-issue.

Code-graph changes

• This PR targets main
• A linked issue agrees on the bounded extractor/resolver scope
• The change follows the frozen LanguageExtractor or FrameworkResolver interface
• A focused fixture and assertions for the expected node/edge shape are included
• Any new grammar WASM, extension mapping, extractor, or resolver is registered
• No graph identity, reconciliation, schema, or drift-semantics changes are included, or a core / discuss-first issue is linked above