bug(native): native engine under-extracts ~68 call edges vs WASM

[carlos-alm]

Description

During v3.4.0 dogfooding (PR #611), the engine comparison revealed that the native engine produces ~68 fewer call edges than WASM (~4,000 vs ~4,068). While the native engine has higher call confidence (80.9% vs 71.3%), the gap suggests either:

1. Native is missing legitimate call sites that WASM extracts, or
2. WASM is over-extracting false call edges at lower confidence

Per CLAUDE.md policy: "If two engines produce different results, that is a bug in the less-accurate engine — not an acceptable parity gap."

Metrics (v3.4.0, codegraph self-analysis)

Metric	Native	WASM	Delta
Total edges	20,849	20,950	-101 (0.5%)
Call edges	~4,000	~4,068	-68 (1.7%)
Call confidence	80.9%	71.3%	+9.6%
Nodes	10,941	10,915	+26 (0.2%)

Next steps

• Investigate which call sites WASM finds that native misses (or vice versa)
• Determine which engine is correct for the divergent edges
• Fix the less-accurate engine

[carlos-alm]

3.4.0 benchmark update (PR #625): The DB size divergence between engines has widened significantly. In 3.3.1 the gap was ~3% (48,707 vs 47,317 bytes/file). In 3.4.0 it's ~37% (55,041 native vs 40,198 WASM bytes/file). Native grew 13% while WASM shrank 15%. This is likely amplified by the new AST-store and complexity phases interacting with the underlying node/edge extraction differences. Worth investigating as part of the parity fix.