perf(native): ~2 second flat overhead added to rebuild operations in v3.10.0

[carlos-alm]

Summary

The v3.10.0 pre-publish benchmark gate detected a catastrophic native-engine regression on rebuild operations. A flat ~2 second overhead was added to every rebuild on the native engine. The publish was correctly blocked from npm by the regression guard (failed run).

Regression deltas (v3.10.0 vs v3.9.6, native engine, 25% threshold)

Build benchmark:

• No-op rebuild: 13ms → 2125ms (+16246%)
• 1-file rebuild: 78ms → 2143ms (+2647%)

Incremental benchmark:

• No-op rebuild: 10ms → 2081ms (+20710%)
• 1-file rebuild: 54ms → 2139ms (+3861%)

WASM engine was unaffected on the same run, so this is native-path-specific.

Pattern

A flat ~2080ms overhead is added regardless of how much work the rebuild has to do (no-op vs 1-file produce nearly identical absolute times). This is the signature of something being initialized or loaded per call that previously persisted across calls — not a slowdown in rebuild logic itself.

Important context

The benchmark CI built v3.10.0 source against the v3.9.6 native binary (the published artifact). Logs show:

[codegraph] Using native engine (v3.9.6)

That means the regression is on the JavaScript/TypeScript wrapper around native, not in the Rust extractor crate.

Suspected causes (PRs between v3.9.6 and v3.10.0)

• perf(wasm): scope ensureWasmTrees re-parse to files that need it #1038 — perf(wasm): scope ensureWasmTrees re-parse to files that need it — touched wasm-worker-entry.ts serialization and ensureWasmTrees filtering. If ensureWasmTrees (or the JS path that wraps native parses) is now invoked on the native path, or the empty-array serialization affects native flow, that would match the symptom.
• fix(parity): gate astTypeMap lookup with Object.hasOwn #1041 — fix(parity): gate astTypeMap lookup with Object.hasOwn — changed extractor lookup behavior. Lower-probability cause but worth ruling out.

Reproduction

git checkout v3.10.0
npm install
npm run build
node --import ./scripts/ts-resolve-loader.js scripts/build-benchmark.ts --version 3.10.0
node --import ./scripts/ts-resolve-loader.js scripts/incremental-benchmark.ts --version 3.10.0

Compare `native.noopRebuildMs` and `native.oneFileRebuildMs` to the v3.9.6 baseline in `generated/benchmarks/{BUILD,INCREMENTAL}-BENCHMARKS.md`.

Acceptance criteria

• Native no-op rebuild back to <50ms (was 10–13ms in v3.9.6)
• Native 1-file rebuild back to <100ms (was 54–78ms in v3.9.6)
• `npm run test:regression-guard` passes against the v3.9.6 baseline
• Pre-publish benchmark gate clears for the next release attempt

[carlos-alm]

Trace results from #1062 — phase breakdown of the 2s native overhead

Dispatched publish.yml against chore/issue-1054-timing-trace to run the gate with per-iteration phase timings. Run: https://github.com/optave/ops-codegraph-tool/actions/runs/25346156742

WASM no-op (fast-skip path fires): 20–31 ms total — every phase shows ? (early exit before phases run).

Native no-op (no skip — full pipeline runs): 2150–2310 ms total. Every phase reports a real number.

phase	native.noop avg ms	share
parse	~620–770	~30%
insert	~400	~18%
ast	~260	~12%
cfg	~205	~9%
edges	~185	~9%
dataflow	~170	~8%
roles	~93	4%
structure/complexity/setup/etc	~150	~7%
sum	~2180	matches total

Diagnosis

There is no single hot phase — the entire pipeline re-runs.

Crucially, native.full[1] = 2380 ms and native.full[2] = 2117 ms (incremental benchmark, second/third iteration) cost the same as native.noop iterations. So full rebuilds after the cold run are also paying the full ~2 s, which means the native engine is not detecting unchanged inputs and is re-parsing/re-analysing every file on every rebuild. WASM's early-skip exit (~20 ms) works correctly; the equivalent fast-path is either missing or unreachable in the native side.

Next step

Focused PR to make the native incremental path skip parse/ast/cfg/dataflow/edges when no source files changed (matching WASM's behaviour). #1062 to be closed unmerged once the fix lands.

Sample trace lines from the run:

[1054-trace] wasm.noop[0]   total=31ms   setup=? collect=? detect=? parse=? insert=? ...   (early-skip)
[1054-trace] native.noop[0] total=2310ms parse=767  insert=406 ast=255 cfg=205 edges=183 dataflow=167 roles=93 ...
[1054-trace] native.noop[1] total=2175ms parse=626  insert=401 ast=278 cfg=206 edges=185 dataflow=169 roles=93 ...
[1054-trace] native.full[1] total=2380ms parse=888  insert=390 ast=249 cfg=198 edges=179 dataflow=168 roles=91 ...