Harden document linter heuristics, report output, and test coverage

[feddernico]

Summary

This PR hardens the Document Linter so its output is more trustworthy, more actionable, and better covered by tests.

What changed

• Removed the fake async setTimeout from the analysis path.
• Unified sentence tokenization behind one helper and documented its limits.
• Expanded markdown parsing to handle:
  • ordered lists
  • Setext headings
  • indented code blocks
  • unterminated fenced code blocks
• Deduplicated pseudo-section-label detection.
• Replaced regex-based “needs attention” inference with explicit section metadata.
• Added an aggregate overall score to both the panel and markdown export.
• Reworked engagement/strength heuristics to use generic signals instead of topic-specific phrases.
• Penalized thin / low-signal drafts more aggressively.
• Added repeated-word detection for obvious low-quality text.
• Extracted linter copy into a dedicated messages.ts module.
• Added a Rerun on change toggle in the linter panel.
• Added clickable section line links that jump back into the editor.
• Preserved panel scroll during refresh and added best-effort focus restoration.
• Made export explicitly announce when it must re-analyze before generating a report.

Tests

Added/expanded QUnit coverage for:

• markdown parsing edge cases
• section building in isolation
• markdown export snapshot output
• empty / heading-only / single-word inputs
• list items not counting as prose sentences
• controller behavior
• low-signal / repetitive draft scoring

Validation

• npm run build:test
• npx qunit tests/qunit/document-linter.test.js
• npm run test:qunit
• npm run build

Notes

• The build still reports two pre-existing ESLint warnings in tests/qunit/auth.test.js.
• The panel still rebuilds its markup on refresh, but scroll is now preserved and focus restoration is attempted, so the refresh is less disruptive than before.