[remigiusz-antczak]

Summary

method.md is the phase procedure that every FILE, FULL and large or flagged DIFF review reads. This PR splits it into a core plus two lazy references (43,524 → 29,936 bytes in the core). The PR builds on the full-map.md split.

• references/method-diff.md (7,872 bytes). A review reads it on a DIFF or FILE scope. It holds:
  • the hunk-kind depth and intent-conformance checks;
  • the count-invariant check;
  • the sweep-scope check;
  • the rewritten browser spec;
  • the fix-completeness sweeps.
• references/method-lenses.md (7,743 bytes). A review reads it when a call site may bypass or re-derive a shared helper, or when a changed unit has no covering test. It holds comment-mining, the canonical-predicate diff and the Phase 3 disposable probe.

A lens cannot fire from a suspicion the agent does not yet have. The core therefore keeps two always-on stubs, each of which points to the full procedure:

• the comment-mining grep, on FULL and FILE reviews of repos that have a shared component dir;
• the canonical-predicate check.

Measurement (offline byte counts; no paid run)

method.md text

Case	Bytes	Change
FULL, no lens fired	29,936	−31.2%
FULL plus lens file	37,679	−13.4%
Large DIFF	37,808	−13.1%
Worst case, both lazy files	45,551	+4.7%

Whole instructed load (SKILL.md + full-map.md + method.md)

Case	Bytes	Change
FULL	64,083 → 50,711	−20.9%
Large DIFF	64,083 → 58,583	−8.6%

The whole-load figures were measured on a full-map.md core of 17,990 bytes. The landed core is 18,654 bytes, because principles 7 and 8 were restored, so each whole-load figure is 664 bytes higher (FULL 51,375, −19.8%; large DIFF 59,247, −7.5%). The typical cut holds only while the lens triggers stay unfired. Recall under the lazy lenses was not measured; a paid quality bench is waiting on owner approval.

Safety

test_method_split.py (8/8) pins:

• that every old line survives (scripts/method-split-lines.txt);
• the reworded and cut lines, each with a mutation check;
• where each lens paragraph lives.

mustload floors are re-pinned to the core. Citations of moved text point to the new files.

Review

• The first independent review found a 🟡 recall risk: the lens triggers fired only on suspicion.
• The fix adds the core grep stubs and objective triggers.
• The re-review, pinned to 80869bb12c69e9c8cd969e06938ef167d3e8240d, found 🔴0 🟡0 🔵1; the nit (a narrower grep in the core stub) is queued.
• This branch is that commit cherry-picked onto main bb8c1cb7ad293f8e16a2c8a33e5b20e2bc7e6e75, which includes the full-map split (Split full-map.md into a lean core plus two lazy references #1953) and the test-flake fix (Stop the temp-repo git gc flake in tests #1952). The table conflicts are resolved: the full-map.md budget is 18,654 with its raise marker, mustload floors are 18,534 and 26,513, and INDEX.md is regenerated.

Tests

• test_method_split 8/8.
• test_fullmap_core_split, the slim and review ci-tests.d scripts, and the local gates pass.
• CI on the exact SHA is the blocking gate.

🤖 Generated with Claude Code