feat(FAFF-915): trim adversarial-review context to diff-relevant regions

[alechill]

Summary

Adds a diff-relevance filter to review-call.mjs that trims the --context bundle to the regions the diff references before the review payload is assembled. Reasoning-ON reviewers empty out on large payloads (FAFF‑906: the model spends its whole output budget reasoning and emits zero findings); the context bundle is most of that payload while the diff only touches a few regions. Trimming it roughly halves the payload so a reasoning-ON review fits under the empty-out knee. This is the targeted context-trim (the lighter fix), not the general decomposer (FAFF‑916, cancelled).

Closes FAFF‑915.

Approach

• A pure, exported trimContextFiles({ contextFiles, diff, ... }) applied between reading the --context files and assembleUserMessage in main.
• Relevance = a diff-adjacency + identifier-window heuristic: keep diff-touched ranges (always) plus lines naming a diff identifier, each expanded by a ±window and merged; drop the rest with an elision marker.
• Gated by a byte threshold, so small payloads are byte-identical to today and every existing review test is unaffected; --context-trim-bytes 0 disables it entirely.
• Conservative: a diff-touched line is never dropped. Identifier anchors are frequency-capped and a retained-fraction ceiling forces head-retention, so a recurring symbol cannot under-trim. Small no-anchor caller-selected files pass through untouched; large no-anchor files retain their head. Fails safe (passthrough) on any diff-parse ambiguity.
• Pure and deterministic, so the trimmed context stays identical across the four spec-review lenses and the shared-prefix cache (FAFF‑903) holds.

Acceptance Criteria

• AC 1 — Above the threshold, a large no-anchor file is head-reduced; a diff-touched file keeps the touched region and its window; a small no-anchor file passes through.
  Verified: test/adversarial-call.test.mjs FAFF‑915 tests 3, 5, 6 — passing
• AC 2 — report.bytesAfter <= 0.6 * report.bytesBefore on the fixed fixture.
  Verified: FAFF‑915 test 4 — passing
• AC 3 — Below the threshold, assembleUserMessage output is byte-identical and the existing suite passes unchanged.
  Verified: FAFF‑915 test 1 + full suite node --test test/adversarial-call.test.mjs → 284 pass
• AC 4 — A line inside a diff-touched range is never dropped.
  Verified: FAFF‑915 tests 7 and "touched over the ceiling keeps all touched lines" — passing
• AC 5 — --context-trim-bytes 0 disables the trim.
  Verified: FAFF‑915 test 2 — passing
• AC 6 — Fail-safe passthrough on a malformed diff.
  Verified: FAFF‑915 test 11 — passing
• AC 7 — Exactly one stderr note when a trim fires, none on the no-op paths.
  Verified: CLI end-to-end — [note] FAFF‑915 context trim: 4800 → 693 context bytes on fire, silent when disabled
• AC 8 — Deterministic across repeated calls (shared-prefix cache holds).
  Verified: FAFF‑915 test 16 — passing
• AC 9 — Frequency-cap and retained-ceiling fallback.
  Verified: FAFF‑915 tests 9, 10 — passing
• AC 10 — faff validate-adapters, unit suite, faff lint-refs clean.
  Verified: validate-adapters PASS (20 slot skills), lint-refs PASS, gate ladder signal pass

Review

Adversarial code review: 4 rounds, terminal verdict pass. Resolved: full-path range matching, blank/CRLF context-line desync, hunk-length-aware parsing, headReduce clamp. Residual findings were non-gating (a refuted false positive, malformed-input-only edges, documented low-risk items).

Closes FAFF-915

[github-actions]

Ready to land. Run this locally in a real terminal:

faff effects declare --run-dir .faff/anchors/run-20260831-141101-graft-FAFF-915 --issue FAFF-915 --step merge <<'EOF'
[{"kind":"merge","target":"pr:793","reversible":true}]
EOF
faff merge-gate --pr 793 --issue FAFF-915 --run-dir .faff/anchors/run-20260831-141101-graft-FAFF-915 --level L2 \
  --execute --merge-args "--squash --delete-branch"