research(bench): benchmark resolution against Jelly and ACG on shared fixtures (Phase 8.6 follow-up)

[carlos-alm]

Context

Phase 8.6 (#1299) deferred the external validation step from the roadmap:

Benchmark against Jelly and ACG on shared fixture projects for external validation

What's needed

1. Select shared fixture languages — Jelly targets JavaScript/TypeScript; ACG targets Java. Pick 2–3 fixture projects from tests/benchmarks/resolution/fixtures/ that overlap (likely javascript, typescript, java).

2. Run Jelly on the TypeScript/JavaScript fixtures and collect its call graph output. Compare edge sets against codegraph's resolved edges and the hand-annotated expected-edges.json manifests. Compute precision/recall for Jelly on the same corpus.

3. Run ACG (or a compatible tool) on the Java fixture similarly.

4. Produce a comparison table — precision, recall, and TP/FP/FN counts for codegraph vs Jelly vs ACG on the shared fixture set. Document in docs/benchmarks/RESOLUTION-COMPARISON.md.

5. Wire into CI (optional) — if Jelly/ACG can be installed in CI without excessive overhead, add a comparison job to the resolution benchmark workflow; otherwise keep it as a manually-run script.

References

• Jelly: https://github.com/cs-au-dk/jelly
• ACG: https://github.com/maccioni/acg (or equivalent)
• Existing fixture format: tests/benchmarks/resolution/fixtures/typescript/expected-edges.json
• Resolution benchmark runner: tests/benchmarks/resolution/resolution-benchmark.test.ts

Notes

• The goal is external validation of codegraph's precision/recall claims, not necessarily matching their numbers — the fixture set is intentionally small and hand-annotated
• Jelly supports whole-program TypeScript analysis; the comparison should use the same fixture source files codegraph builds against