Harden Compass performance and improve semantic graph dominance

[forhappy]

Summary

• add a correctness-gated performance harness for cold, warm, incremental, natural-language query, and CompassQL workloads
• verify Compass and Graphify on pinned Django (Python/JavaScript) and Entire (Go/shell) repositories
• preserve path-sensitive cache identity, sealed unchanged-build reuse, empty-file inventory, and deterministic byte-stable publication
• resolve Python imported types through re-exports and retain safely qualified unresolved external member calls
• extract bounded JavaScript prototype / .fn methods and correctly anchor grouped/generated Go types
• classify cross-schema facts as exact, semantically dominated, ambiguous, or missing without label-only matching
• reduce publication work through owned fact transfer, canonical ordering, bounded anchor parsing, deferred failure diagnostics, and indexed Python call sites
• improve identifier-aware query ranking, including CamelCase symbols such as URLResolver

Why

Compass needed repeatable, fail-closed evidence for both performance and graph quality. Literal Graphify comparison mixed real omissions with cases where Compass owns a stronger canonical definition. The implementation therefore improves native extraction/resolution first, then uses Graphify only as a development oracle to prove exact coverage or bounded semantic dominance.

Root causes fixed in this phase include lexical/canonical cache-root disagreement, empty generic files being cleared after extraction, Python re-export provenance stopping short of anchored definitions, qualified external calls losing their endpoints, JavaScript prototype assignments not becoming owned methods, and Go receiver placeholders winning over later real or grouped type declarations.

Real-repository quality results

Both final graphs have zero validation errors.

Repository	Graphify fact	Exact	Dominated	Ambiguous	Missing	Exact + dominated
Django	Nodes	50,466	6	25	348	99.27%
Django	Edges	149,993	336	26	8,355	94.72%
Entire	Nodes	19,866	663	6	50	99.73%
Entire	Edges	53,593	6,104	18	1,347	97.76%

This improves substantially over literal missing counts of Django 1,175 nodes / 10,600 edges and Entire 720 nodes / 7,470 edges. Strict Graphify-superset quality is intentionally not claimed: the remaining cases fail closed and are concentrated in config/vendored-resource relations, shell entry points, Go embedding occurrences, and ambiguous inherited/generated identities.

The final generated-Go correction publishes ErrorModelStatusCode at its declaration and optionFunc as an anchored type_alias; on Entire it reduces ambiguous nodes from 20 to 6 and missing nodes from 52 to 50.

Performance

Pinned corpora: Django 50d706d0, Entire 279b9885. Fresh comparison: Graphify 0.9.31 at 4fe11092; historical baseline: Graphify 0.9.30 at ecfcd160.

Tool	Repository	Workload	p50	p95	Peak RSS
Compass	Django	Cold, final 3	12.355s	16.587s	4.45 GiB
Compass	Django	Warm, final 5	2.182s	2.429s	674.58 MiB
Graphify 0.9.31	Django	Cold, 3	49.985s	50.916s	1.26 GiB
Compass	Entire	Cold, 3	4.055s	4.127s	1.42 GiB
Compass	Entire	Warm, 3	0.586s	0.599s	247.25 MiB
Graphify 0.9.31	Entire	Cold, 3	17.650s	18.214s	226.09 MiB

Django remains 5.05x faster than the pinned Graphify 0.9.30 median. Graphify 0.9.31 itself became materially faster, so current-head cold ratios are 4.05x on Django and 4.35x on Entire. Django warm is 11.96% above the prior 1.949s baseline, narrowly outside the 10% phase gate. Peak RSS also remains higher than Graphify. These remain explicit follow-up performance gaps.

Query verification

Single fresh-process semantic checks (not latency medians): Compass passed all four repository oracles. Graphify passed three; its Django URL-resolution result omitted required URLResolver evidence.

Query	Compass	Graphify 0.9.31
Django URL resolution	2.783s, pass	2.414s, fail oracle
Django model save	3.303s, pass	4.232s, pass
Entire checkpoint creation	0.655s, pass	1.199s, pass
Entire repository state	1.614s, pass	1.923s, pass

Validation

• cargo fmt --all -- --check
• cargo test -p compass-files --test contracts (23 passed)
• complete compass-languages, compass-resolve, and compass-graph suites
• graph-v1 determinism (8 passed) and publication resilience (10 passed)
• python3 -m unittest discover -s benchmarks/performance/tests -v (60 passed)
• release build and fresh Django/Entire graph builds with zero validation errors
• git diff --check
• parent graphify update . refresh completed

Detailed implementation context, exact constraints, methodology, and remaining gaps are in docs/superpowers/plans/2026-07-30-semantic-graphify-dominance.md and docs/superpowers/reviews/2026-07-30-compass-performance-baseline.md.