Epic: compiler call-graph comparison across languages (tooling, skills, per-language references) #3772

Goal

Measure and improve GitGalaxy's function call graph language by language, against a compiler or type-checker reference. Today there are two references: pyan3 for Python (#3332/#3378) and the TypeScript type checker (#3761). Each new language should be cheap in time and tokens.

What the TypeScript round taught us

The TypeScript round (#3756–#3760, #3764, #3767) was spent mostly on five things:

1. Building the reference. Several iterations, and TypeScript 7 shipping no JS API broke CI.
2. Triaging misses and wrong links by hand. Throwaway scripts, then reading samples by eye. This was the biggest token cost.
3. Environment detours. Missing test and ML dependencies, a different ruff version, and the tokenizer download blocked by the network policy.
4. The same verification chain run by hand on every PR. About six commands, each with long output.
5. Fixes forced to land one at a time.

Plan

Infrastructure: one PR

• Call-graph comparison: a reference-adapter contract, with a cached reference graph #3773 A reference-adapter contract, with a cached reference graph
• callgraph_triage.py: bucket every missed and wrong call link into a ranked backlog #3774 callgraph_triage.py: miss and wrong-link buckets as a ranked backlog
• callgraph_check.py: the call-graph verification chain as one command with a short summary #3775 callgraph_check.py: the whole verification chain as one command with a short summary
• Call resolver: explain one link (rung, candidates, qualifier) without reading call_resolver.py #3776 Resolver explain for a single link
• Session setup hook: install the pinned call-graph toolchain for Claude Code on the web #3778 A session setup hook that installs the pinned toolchain

Process

• Skill: callgraph-language-onboarding (add a compiler reference for a new language) #3779 The callgraph-language-onboarding skill (checklist)
• Skill: callgraph-miss-sweep, plus a Haiku callgraph-triage-scout agent #3780 The callgraph-miss-sweep skill, plus a Haiku callgraph-triage-scout agent

Languages (rough order)

• Call-graph reference: JavaScript (TypeScript checker with allowJs) #3781 JavaScript: the TS adapter with allowJs
• Call-graph reference: Go (golang.org/x/tools static call graph) #3782 Go: golang.org/x/tools static call graph
• Call-graph reference: evaluate SCIP on Java as a shared route for several languages #3783 Java: evaluate SCIP as a shared route for several languages
• Call-graph reference: C/C++ (clang-based) #3784 C/C++: clang-based
• Call-graph reference: Rust (rust-analyzer SCIP) #3785 Rust: rust-analyzer SCIP

Still open: whether to adopt SCIP (#3783), and whether fixes can run in parallel on more than one branch.

Ground rules

• A reference is a reference, not ground truth. Disagreements are checked against source before they're claimed (CLAUDE.md, "Comparative-correctness claims").

• Every finding is filed in this milestone ("Compiler call-graph comparison").

• Cross-language patterns are expected to recur:

  • bodyless interface/abstract members (Call resolver (TypeScript): confident links land on bodyless signatures (interface / abstract / type members) instead of the code that runs #3757's pattern) in Java, C#, Go and Rust;
  • calls inside string interpolation;
  • built-in method shadowing (Call resolver: an untyped receiver's built-in method call (str.trim(), map.get()) links confidently to a same-named repo method via file/import #3756).

  Triage should name these buckets, so a known pattern is recognised rather than rediscovered.