perf: hash tx effects tree nodes synchronously

[vezenovm]

Summary

Switches only txEffectsTreeNodeHash to the existing synchronous Poseidon helper through the initialized crypto/sync entrypoint.

• preserves the Promise-returning hasher interface, TX_EFFECTS_TREE domain separator, input serialization, and output
• leaves effect-category and tx-effect leaf hashing asynchronous
• also accelerates membership construction and verification, which share this internal-node hasher
• keeps the existing backend policy: native shared memory where available, synchronous WASM fallback elsewhere
• adds no transport setting, bb endpoint, batching, or general tree refactor

Benchmark methodology and the reusable harness are in #260. The harness files and raw output are intentionally absent from this PR.

Same-revision transport benchmark

Measured on this PR's origin/main base and toolchain in two focused runs: old production UDS hashing versus UDS leaves with the new sync-SHM internal hasher. Each scenario reports the median of 15 samples after three warmups. Times are milliseconds per complete operation.

Profile	Txs	Before	After	Change
Sparse	1	0.135	0.135	inconclusive
Sparse	3	0.292	0.271	−7.3%
Sparse	16	1.510	1.317	−12.8%
Sparse	64	6.116	5.395	−11.8%
Small categories	1	0.189	0.189	inconclusive
Small categories	3	0.433	0.429	inconclusive
Small categories	16	2.319	2.193	−5.4%
Small categories	64	10.630	9.833	−7.5%
Log-heavy	1	8.784	8.531	−2.9%
Log-heavy	3	26.400	25.316	−4.1%
Log-heavy	16	141.561	138.267	−2.3%
Log-heavy	64	574.766	575.820	inconclusive

The earlier 5–8% expectation was a hypothesis. The measured complete-commitment result is workload- and tree-size-dependent: 7.5% for the 64-transaction small-category case, 11.8% for sparse/64, and no detectable change when large leaf hashes dominate.

Internal tree and membership

These 64-leaf cases use precomputed leaves, isolating the shared internal-node path.

Workload	Before ms	After ms	Change
Internal tree	1.839	0.947	−48.5%
Membership construction	1.937	1.050	−45.8%
Membership verification	0.178	0.092	−48.6%

Checkpoint benchmark

Before and after runs used the same current-main revision and toolchain. Each case has three warmups and 15 measured samples; every sample creates a fresh builder and world-state fork. Mixed direction and small differences make these checkpoint-level results inconclusive.

Profile	Txs	buildHeaderAndBody before	after	Total before	after
Worst-case	4	27.677	25.899	40.441	38.243
Worst-case	8	53.111	51.788	76.758	75.188
Worst-case	16	108.255	105.194	155.529	151.816
Class-log-heavy	4	57.112	59.838	82.731	87.643
Class-log-heavy	8	114.993	112.686	165.362	164.033

Responsiveness

Median of three saturated probe rounds with four concurrent 64-transaction small-category commitment loops. This is synthetic saturation, not a node RPC latency prediction.

Metric	Before	After
Complete ops/s	96.8	93.1
Worker ping p50	0.05 ms	0.05 ms
Worker ping p95	13.66 ms	5.35 ms
Worker ping max	16.18 ms	12.84 ms
Main timer max delay	11.94 ms	10.19 ms

Throughput and latency differences in this probe are within run-to-run noise; importantly, selectively synchronous internal hashing does not reproduce the roughly one-second starvation seen when all category and leaf hashing is synchronous.

Validation

• ./bootstrap.sh from yarn-project passed after updating to current main.
• Full TypeScript build passed.
• Repository-wide format and lint checks passed.
• Body and tx-effect membership suites passed before and after: 18 tests covering empty, single-leaf, balanced, unbalanced, and invalid-witness cases.
• The same 18 tests passed with BB_BINARY_PATH forced to a nonexistent path, exercising the synchronous WASM fallback.
• Focused before/after transport benchmark runs and both 15-sample checkpoint runs passed.

Raw samples remain in ignored local output. Blob accumulator batching, block-header trees, and the full sequencer benchmark are out of scope.