feat(review): filter unreviewable content from diff context

[guyoron1]

Review context includes lockfile, minified, sourcemap, and @generated
hunks that no model can meaningfully assess; on a mixed PR they are
pure input cost for all dimensions. This filters them deterministically
before context assembly — migrations exempt, the exclusion always
disclosed in the review body so a changed lockfile is still visible.
Small-PR and large-PR paths now share one definition of "generated"
instead of a prompt-level list.

What changed

• New skills/pr-review/scripts/filter-review-diff.sh: a single awk
  program over a unified diff on stdin. Strips lockfile sections
  (mirrors the is_lock() regex in fullsend's
  .github/scripts/route-review-model.sh, comment cites the source),
  *.min.js/*.min.css, *.map, and vendor//node_modules//
  third_party/ paths, plus any section with @generated in its first
  5 added lines. migrations//migrate/ paths are exempt from every
  rule. Filtered diff on stdout (byte-identical for untouched
  sections), one-line-per-file exclusion summary written only to the
  file given as $1, never to stdout.
• scripts/filter-review-diff-test.sh, wired into make script-test:
  byte-identical pass-through (via cmp), lockfile/mixed-diff
  stripping, the added-vs-removed-line @generated distinction, the
  migration exemption, minified/sourcemap/vendored stripping, summary
  format and stdout-silence, all-stripped output, and malformed-input
  pass-through. All pass locally (bash 3.2, macOS); shellcheck clean.
• skills/pr-review/SKILL.md step 2: small-PR path now pipes the full
  diff through the filter before it enters any context package; the
  large-PR per-file path's "lockfiles, vendor/, protobuf, etc." vibes
  list is replaced with the same script, so both paths share one
  deterministic definition. FILE_COUNT/LINE_COUNT stay computed
  from the unfiltered file-stats response (unchanged) — routing must
  see the true size.
• skills/pr-review/SKILL.md step 3d's diff context-package
  description updated to match.

Design note: disclosure via info finding, not a footer

An earlier draft disclosed the exclusion through a review-body footer.
Step 7 forbids that explicitly ("No footer. Do not append any
footer, action-hints block, or boilerplate after findings."), so the
disclosure uses the file's existing precedent for orchestrator-level
process disclosure instead: an info-level finding, the same mechanism
the provenance-warning finding (right above it in step 7) already
uses. The new excluded-content info finding sits alongside it.

Rebase note

Open PR agents#1006 (feat/review-untrusted-text) also touches
SKILL.md — its hunks are in step 6d's Part 3/4 blocks, mine are in
steps 2 and 3d. Different sections, shouldn't collide, but noting the
adjacency. #1004's step 6g hunk is further down in step 6, also should
not overlap.

Non-goals

No filtering of step 2b's file-contents fetch beyond what already
exists there, no change to route-review-model.sh, no config knob.

Scope

The filter script is architecture-neutral; the wiring here is the claude
review path's orchestrator (skills/pr-review). The pi-runtime review
(no sub-agents) assembles its own context and can adopt the same script
when that path stabilizes.

[github-actions]

Functional tests did not run

Functional tests run automatically for org/repo members and collaborators on pull requests.

For other contributors, a maintainer must add the ok-to-test label after the latest push.