fix(decision-records): respect code spans, validate link anchors

[trungfinity]

Problem

check_links had two gaps:

1. False positives on code spans. It regex-scanned raw text, so a link written inside a code span or fenced block — syntax being displayed, not a live link (per CommonMark, code spans contain no links) — was flagged as broken. Hit in practice by an AGENTS.md line showing the link format as an example: [term](../../glossary.md#term).
2. Anchors never validated. The fragment was discarded (target.split("#", 1)[0]), so glossary.md#egres-proxy passed as long as glossary.md existed, and pure #fragment intra-doc links were skipped entirely. Repos adopting glossary-term and section-heading linking get zero protection on the half of the link most likely to rot (promotion moves records between directory depths).

Fix

• strip_code() blanks fenced blocks and inline code spans before the link scan (line structure preserved).
• heading_anchors() computes GitHub-style slugs for a file's headings — lowercase, punctuation dropped, spaces→hyphens, -N suffixes for duplicate headings, markdown links/backticks in heading text reduced to display text (covers em-dash cases like #appendix--harnesses-evaluated-research-notes-non-binding).
• Every #fragment on a markdown target is validated against the target's slugs; pure #fragment links resolve against the containing file. Failures report as broken anchor, distinct from broken link.

Testing

• 5 new tests: slug computation (em-dash, quotes, &, duplicates, fenced pseudo-headings), code-span and fenced-block links not checked, cross-file anchor validation (pass + fail), intra-doc anchor validation (pass + fail).
• Full suite: 62 passed.
• Field-tested against skymavis/lanh-ai's docs tree (~70 glossary/section anchor links across 6 accepted records + 3 drafts): all validate; the previously false-positive AGENTS.md pattern no longer flags.