docs: publish verified Graphify native-memory comparison

[suhaanthayyil]

Summary

• publish the verified full Graphify/Entire Graph/CMM native-memory comparison in the README
• add the frozen revisions, fairness boundary, integrity record, and external evidence link to benchmark docs
• state the split result honestly: Entire Graph wins LOCOMO retrieval recall; Graphify wins final-answer accuracy
• skip the max-RSS guard assertion on Windows, where the benchmark explicitly reports RSS as unavailable; production code and the benchmarked binary are unchanged

Result

• LOCOMO recall@10: Entire Graph 0.918, Graphify 0.844
• LOCOMO QA: Entire Graph 51.0%, Graphify 67.4%
• LongMemEval-S QA: Entire Graph 44.7%, Graphify 62.7%
• graph-build LLM credits: 0 for Entire Graph, Graphify, and CMM

Boundary

This PR is stacked on the exact benchmarked candidate commit dea450b. Its only non-documentation change is a test-only Windows skip for a metric that rss_windows.go deliberately cannot measure. It does not change the measured binary. The comparison is explicitly labeled a public-protocol reimplementation because Graphify does not publish its advertised memory harness or selectors.

The protocol, 1%/10%/100% sequence, fairness audit, and artifact hashes are under review in entirehq/graphmark#63.

Validation

• go vet ./...: passed (upstream tree-sitter Lua compiler warning only)
• go test ./...: passed locally
• git diff --check: passed
• prior Windows CI failure was isolated to the unsupported RSS assertion; follow-up CI is the cross-platform confirmation

Stacked on #77. Review-only: do not merge until both benchmark PRs are approved.

[suhaanthayyil]

Superseded by #84, the single Entire Graph release PR now targeting main. The older PR is preserved for history; its branch is not deleted.