feat(benchmark): call-graph focused benchmark + tree-sitter end_line fix

[jafreck]

Summary

Two related improvements: a benchmark overhaul focused on call-graph questions, and an indexer fix that patches SCIP symbol end_line using tree-sitter data.

Benchmark Overhaul

Question catalog: Replaced the 15-question mixed catalog (call graph, type hierarchy, module deps, test mapping, complexity, API surface, architecture) with 13 call-graph focused questions where 10 leverage lore_snippet for code-level understanding.

Key changes:

• Added symbol2 placeholder for Q1.5 (shared callers) and Q1.7 (call chain trace)
• Redesigned Q1.7 from "trace deepest chain" (nondeterministic) to "trace from X to Y" (deterministic)
• Redesigned Q10.1 from "rank top 3 by fan-in" (nondeterministic) to "list all callers of X" (deterministic)
• Fixed incorrect ground truth across all repos (verified against pinned SHA DBs)
• Removed prose from expectedAnswer; use only concrete tokens (function names, file paths) for reliable substring scoring

Ground truth fixes:

• esbuild Q1.1: callers of Build are runImpl/handleBuildRequest, not main/rebuildImpl
• jackson-databind Q7.3: createCollectionDeserializer delegates to findTypeDeserializer etc., not sibling factories
• lore-self Q7.3: resolutionStage only calls resolveSymbolEdges (no resolveImports)
• lore-self Q10.1: top fan-in is setLoreMeta(5 files), not openDb
• fastapi Q1.2/Q7.3: add_api_route calls get_value_or_default/APIRoute, not get_request_handler
• jackson-databind Q1.2: strip receiver qualifiers (type.getContentType → getContentType)

Indexer Fix: Tree-sitter Patches SCIP Symbol end_line

Problem: scip-go provides empty enclosingRange for all definitions. The SCIP stage's estimateSymbolEndLine heuristic fails for Go methods with complex parameter types (e.g., map[string]interface{}) because naive brace-counting sees {} in the type annotation.

Fix: SourceIndexStage.computeMetricsForScipFiles (which already parses SCIP-sourced files with tree-sitter for metrics) now also patches end_line when tree-sitter provides a wider range. Includes receiver-prefix name matching (Server.HandleRequest → HandleRequest).

Impact: Go method handleBuildRequest goes from end_line=596 (wrong, just signature) to end_line=790 (correct, full body). This unlocks 20 outgoing call edges and makes cross-file callers like handleBuildRequest → Build visible.

Test: Integration test creates a Go method with map[string]interface{} params, a synthetic SCIP index with no enclosingRange, runs both stages, and asserts tree-sitter corrects end_line.

Benchmark Results (Run 5)

Repo	Lore Correctness	Pass Rate
lore-self	98.7%	14/14
jackson-databind	96.2%	13/14
postgres	85.6%	12/14
zod	83.7%	14/14
fastapi	73.7%	14/14
esbuild	59.4%	14/14

Average lore-arm correctness: ~83% (up from ~27% at session start).

[codecov]

Codecov Report

❌ Patch coverage is 81.25000% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 87.82%. Comparing base (508f9be) to head (1ad94db).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/indexer/stages/scip-indexer.ts	40.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #219      +/-   ##
==========================================
+ Coverage   87.49%   87.82%   +0.33%     
==========================================
  Files          85       85              
  Lines        9475     9482       +7     
  Branches     2932     2936       +4     
==========================================
+ Hits         8290     8328      +38     
+ Misses       1185     1154      -31     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.