Preserve source-backed graph semantics and accelerate publication

[forhappy]

Summary

• publish Bash entrypoints and Go embeddings as first-class typed graph facts
• replace broad Python import-to-class inference with exact decorator occurrences and qualified re-export resolution
• preserve package-qualified Go type identity and prevent external types from binding unrelated local names
• distinguish exact, dominated, rejected, ambiguous, and genuinely missing Graphify evidence
• accelerate deterministic publication with indexed lookups, single proven-clean validation, and ordered parallel preparation
• define the next universal source-grounded quality program and its independent precision/recall gates

Final real-corpus verification

Repository	Compass cold p50	Graphify 0.9.31 cold p50	Speedup	Compass warm p50	Compass query p50 range
Django	9.3868s	49.985s	5.33x	1.6324s	1.4484-1.4598s
Entire	3.5001s	17.650s	5.04x	0.5391s	0.7060-0.7148s

• all five cold samples per repository were eligible with identical correctness digests
• all 40 measured semantic query samples passed their oracles
• both final graphs index with zero validation errors
• Django: 55,120 Compass nodes / 130,847 distinct edges; 99.25% node and 95.84% source-grounded edge coverage of the retained Graphify baseline
• Entire: 21,711 Compass nodes / 72,250 edges; 99.77% node and 98.50% source-grounded edge coverage, with 11,188 more edges than Graphify
• final Compass verification reused retained Graphify 0.9.31 artifacts; Graphify was not rerun as a performance benchmark

Strict literal Graphify-superset coverage is not claimed. The comparator classifies 53,005 Django edges whose reported owner-to-target evidence is an import occurrence. That occurrence proves import availability, not that every projected owner uses the target; some dependencies may still be real at later occurrences and must be recovered separately or recorded as Compass omissions. On Entire, 2,889 baseline edges conflict with qualified external targets at the exact source site, including audited same-name local rebinding examples. Residual ambiguous and missing facts remain fail-closed and visible.

Universal quality follow-up

The checked-in design defines:

• a versioned universal declaration, scope, binding, occurrence, candidate, and resolution-evidence model
• registered language adapters with explicit complete or legacy capability profiles
• a shared constrained resolver with no repository-wide terminal-label fallback
• framework evidence packs that cannot invent occurrences or bypass resolution constraints
• a deterministic audit of at least 2,000 relationships
• at least 99.5% observed overall precision with a two-sided 95% Wilson lower bound of at least 99%
• at least 99% precision and 95% source-derived recall for every advertised capability
• zero tolerance for fabricated occurrences, cross-language matches, and unsafe local-target substitution

Verification

• cargo fmt --all -- --check
• cargo test --workspace --all-targets
• focused graph publication, language resolution, model validation, and comparator suites
• five-sample release cold/warm/incremental matrices on pinned Django and Entire
• ten measured fresh-process runs for each of four semantic query oracles
• graphify update . from /Users/haipingfu/graphify (no topology changes)