bug: native engine silently drops 60 files vs WASM in 3.9.4 benchmark (engine parity regression)

[carlos-alm]

Summary

In the v3.9.4 build benchmark (PR #958), the native engine processed 668 files while the WASM engine processed 728 files — a 60-file (~8%) gap. In v3.9.3 both engines ran on the same 727 files, so something regressed between 3.9.3 and 3.9.4.

Flagged by Greptile: #958 (comment)

Why this matters

CLAUDE.md states:

Dual-engine architecture: [...] Both engines must produce identical results. If they diverge, the less-accurate engine has a bug — fix it, don't document the gap.

A divergence in the set of files processed (not just results per file) means the native engine is silently failing to collect or parse ~60 files that WASM handles fine. Consequences:

• Per-file metrics (build ms/file, nodes/file, edges/file) are computed from different denominators — engine-vs-engine comparisons in the benchmark table are misleading.
• Anything the native engine silently drops is missing from downstream graph output in production use.

Evidence

From generated/benchmarks/BUILD-BENCHMARKS.md in PR #958:

Version	Native files	WASM files
3.9.2	727	727
3.9.3	727	727
3.9.4	668	728

Investigation hints

• File collection / ignore logic: src/domain/graph/builder/ — check whether the native path now applies a different filter than WASM.
• Native addon: crates/codegraph-core/ — silent parse failure could be swallowed rather than falling back to WASM per file.
• Recent changes touching the native engine or file collection between 3.9.3 and 3.9.4 tags.

Acceptance

• Native and WASM process the same file set (modulo parsers not available in one engine, which must be explicit, not silent).
• A CI check or benchmark assertion fails if the file counts diverge by more than a known, documented delta.