feat: Improve Java/Python resolution, context relevance, and multi-symbol aggregation

[colbymchenry]

Summary

• Eliminate cross-language false positives in name resolution for multi-language projects (Python+Rust, Python+Java, etc.)
• Deprioritize test files in context building so codegraph_context returns production code first
• Improve Java extraction — proper method invocation parsing, interface inheritance, import resolution
• Aggregate results across all matching symbols in callers/callees/impact tools
• Impact analysis traverses into container children so class impact includes method callers

Benchmark

Tested on a 96-file Java project and a 115-file Python+Rust project.

Java: CodeGraph delivered production-quality results across all benchmarks — symbol search (10x richer than grep), call graph (17 callees in 1 call vs impossible with grep), impact analysis (transitive chains), and context building (entry points + code in 1 call).

Python (before fix): 37% of all edges were false positives from Python built-in methods resolving to Rust functions. codegraph_context returned test functions instead of production code.

Python (after fix):

Metric	Before	After
Total edges	4,717	4,640
False cross-language edges	~1,752	28 (mostly subprocess.run → Rust run)
submit_message callees	6 (included Rust extend ❌)	4 (all correct Python ✅)
QueryEnginePort impact	Mixed Python + Rust noise	Clean Python-only results

What changed

Python resolution fixes (index-time) — src/resolution/

name-matcher.ts:

• Added language boundary checks to all 3 matchMethodCall strategies — cross-language class/method matches are skipped
• findBestMatch: cross-language penalty of -80 points (was 0, only had +50 same-language bonus)
• matchByExactName: single cross-language matches get 0.5 confidence (was 0.9)
• matchFuzzy: prefers same-language candidates; cross-language get 0.3 confidence (was 0.5)

index.ts (isBuiltInOrExternal):

• Filters Python built-in type method calls (list.extend, dict.update, str.split, etc.)
• Filters bare Python built-in method names (append, extend, pop, keys, values, join, etc.)

Context relevance fixes (query-time) — src/context/, src/search/

• New isTestFile() utility — detects test files across Python, JS/TS, Go, Rust, Java conventions
• scorePathRelevance() applies -15 penalty to test files (unless query is about tests)
• Context builder reduces test file scores to 30% after result merging
• Re-sorts results so production code surfaces as entry points first

Java extraction improvements — src/extraction/tree-sitter.ts

• Handle Java method_invocation AST node (receiver.method() pattern via object+name fields)
• Support extends_interfaces and super_interfaces with type_list wrapper for Java
• Create unresolved references for Java imports enabling cross-file resolution
• Extract interface inheritance via extractInheritance

Multi-symbol aggregation — src/mcp/tools.ts

• codegraph_callers, codegraph_callees, codegraph_impact now aggregate results across ALL matching symbols (e.g., multiple overloads, same-named methods in different classes)
• New findAllSymbols() helper with note when results span multiple symbols

Impact traversal — src/graph/traversal.ts

• Impact analysis traverses into container children (class → methods) so callers of contained methods appear in the impact radius of their parent class/interface

Other

• deleteSpecificResolvedReferences() in src/db/queries.ts for precise cleanup after resolution
• Added 'instance-method' to resolvedBy union type in src/resolution/types.ts
• Version bump to 0.6.8

Test plan

• All 395 previously passing tests still pass
• Resolution tests: 16/16 pass
• Context tests: 17/17 pass
• Verified via direct DB query: no false cross-language extend/append edges after re-index
• Java benchmark: all 5 categories (search, call graph, impact, context, files) validated
• Re-run Python benchmark after MCP server restart to verify context relevance improvement end-to-end

🤖 Generated with Claude Code