Move diff parsing, three dedup passes, the confidence/severity gate,
inline-eligibility snapping, and payload + fallback rendering out of
skill prose and into ~/.claude/bin/code-review-helper. Cuts the skill
body roughly in half (-156 / +33 net) and makes every deterministic
rule unit-tested + golden-tested against three real OSS PR fixtures
(vercel/next.js#93491, kubernetes/kubernetes#138754, prisma/prisma
#29514).

Capturing those fixtures uncovered three accuracy bugs in the dedup
logic, all fixed in this commit with regression tests:

- Semantic dedup duplicate-emit: in[i] = keep mutation left the
  surviving finding at two slice slots, so the final loop emitted it
  twice. Replaced the in-place rewrite with a current-by-ID map +
  preserved-order slice.
- Cross-ref cascade: positional dedup mutated Explanation to embed
  the dropped peer's file path, which then tripped semantic Rule 1
  on any later finding in that file. Cross-refs now live on a new
  Finding.CrossRefs []CrossRef field rendered at output time; the
  matchers always see the specialist's pristine Explanation.
- Non-deterministic positional sort tiebreak when conf, domain, and
  specialist all matched. Added a final cluster[i].ID < cluster[j].ID
  comparator.

Settings hook now routes .go files through gofmt and other extensions
through prettier, and tools/code-review-helper/bin/ is gitignored.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>