perf: investigate 1-file incremental rebuild regression in 3.4.0

[carlos-alm]

Context

Benchmark PR #627 recorded a disproportionate regression in 1-file incremental rebuild times in v3.4.0:

• Native: 346ms → 448ms (+29%)
• WASM: 508ms → 617ms (+21%)

The file count only grew ~7% (442 → 473 files), so the regression is disproportionate to codebase growth. For a single-file incremental rebuild, total file count should have minimal impact on latency.

Additionally, the WASM full build regressed +15% (4.4s → 5.0s), widening the native/WASM gap from ~1.85x to ~2.21x.

Investigation areas

• Incremental rebuild path in domain/graph/builder.ts
• Hash comparison and file-skip logic
• Any new per-file overhead introduced in 3.4.0

References

• Benchmark data: docs: update incremental benchmarks (3.4.0) #627
• Greptile review comment flagging the regression

[carlos-alm]

Investigation results

The ~100ms incremental regression (native: 346→448ms, WASM: 508→617ms) has been traced to three root causes, all already fixed in post-3.4.0 PRs:

Cause	Est. impact	Fix
classifyNodeRoles ran on ALL nodes every incremental build	~50-80ms (was 255ms at scale)	#622 — incremental roles, 255ms→9ms
Advisory queries in finalize ran unconditionally	~40-60ms	#622 — skipped for incremental builds
Missing compound indexes on edges table	~30-50ms	#632 — added idx_edges_kind_target/source

The TS migration (3.4.0) grew the graph ~32% (more nodes/edges from .ts files), which amplified these O(N) operations disproportionately.

Remaining minor bottlenecks (~10-20ms each, not actionable)

• O(N) file stat in detectChanges: mtimeAndHashTiers() stats all files when journal tier doesn't apply. Inherent — need to detect changes somewhere.
• O(N) directory walk in collectFiles: Full readdirSync walk runs before change detection. Required to detect new/removed files.
• O(N) node loading in buildEdges: Loads all ~10.9k nodes for target resolution. Fundamentally needed — a call in the changed file could reference any function.

These are inherent pipeline costs and not worth optimizing at this scale. The major regression is resolved.

Recommendation: Close this issue. The next benchmark run should confirm the fix.

[carlos-alm]

Closing — all three root causes fixed in PRs #622 and #632 (already merged). See investigation above.