perf(resource): reduce EIP search CPU by 93%

[Savid]

Summary

Optimize the per-query hot path in EIPIndex.Search while preserving the same ranked EIPs and scores.

The cold, held-out confirmation improved ns_per_op from 1,184,777 to 78,539 ± 463: approximately 93.37% less CPU time / 15.1× throughput for the measured search workload.

What changed

• Precompute lowercase EIP title, description, and content once in NewEIPIndex instead of lowercasing the corpus for every query.
• Build a conservative 4096-bit character-trigram Bloom filter for each EIP. Failed probes skip the exact substring scan; successful probes still run strings.Contains, so Bloom collisions cannot change results.
• Precompute query trigram probes, order them rarest-first, and use a build-time inverted index to visit only EIPs that contain the rarest probe bit.
• Convert chunk embeddings to float64 once at index construction and convert each query vector once per search.
• Inline the four-accumulator dot product in the chunk-scoring loop.
• Replace the per-query dedup map and intermediate chunk-score allocation with a dense best-score slice.
• Replace the full result sort with bounded top-k heap selection, followed by a typed descending sort of the selected results.
• Gate the EIP-number regexp behind a cheap case-insensitive eip byte scan.
• Derive result membership from the positive best-score invariant, avoiding the redundant presence slice used during optimization.

There is no public API or wire-format change.

Benchmark result

The benchmark times only EIPIndex.Search over freshly seeded, all-distinct queries. The baseline used three runs after a discarded warmup and measured 1,184,777 ns/op at approximately ±0.28% jitter. Optimization rounds used five interleaved candidate/incumbent pairs on the same VM, with a fresh seed per round.

The final artifact was then cold-booted and measured with held-out seed 1020, which was never used for selection:

Measurement	ns/op	Reduction vs baseline
Baseline	1,184,777	—
Loop best	78,692	93.36%
Cold held-out confirmation	78,539 ± 463	93.37%

Best-so-far progression

Per-round paired evidence

Correctness and anti-gaming controls

The experiment's frozen oracle checked:

• identical ranked EIP IDs and scores within 1e-6;
• concurrent use under the race detector;
• input immutability;
• no returned-result aliasing;
• the repository's existing test suite;
• changing benchmark seeds, preventing a lookup table or memo cache from winning on replayed inputs.

The promoted source initially matched the cold-confirmed snapshot byte-for-byte. The only post-promotion cleanup was deleting the now-unused predecessor filterMayContain helper after the precomputed-probe implementation replaced every call; this has no runtime behavior.

Evidence caveats

The final artifact did reproduce cold, so the end-to-end result is the number to trust. However, the experiment's accepted-gain ledger did not reconcile: it over-counted gains by 134,661 ns because round 8 compared against a stale pre-round-7 incumbent backup and re-banked most of the round-7 Bloom-filter gain. Consequently, this PR does not claim that the individual round gains add up; it claims only the independently confirmed baseline-to-final result.

Round 20's removal of the presence slice was retained as a simplicity tiebreak. Its isolated paired gain, 352.8 ± 198.8 ns, did not clear the configured 556.6 ns acceptance threshold. The aggregate artifact containing that change nevertheless passed the held-out cold confirmation and correctness oracle.

Tradeoffs

• Index construction now does additional lowercasing, trigram indexing, and float conversion work.
• The index retains precomputed lowercase text, Bloom filters, an inverted index, and float64 chunk vectors, increasing resident memory in exchange for substantially lower per-query CPU.
• The benchmark targets steady-state search latency; initialization time and memory growth were not benchmarked by this experiment.

Validation

• go test ./...
• go test -race ./pkg/resource/...
• golangci-lint run ./pkg/resource/...

All pass locally.

[github-actions]

🐼 Smoke eval — c99b2b6: ✅ 8/8 pass

📊 Interactive report — tokens p50 16,595 · tokens/solve 19,498.

Reference points: master@a045b91 100% · v0.38.1 100%.

question	result	tokens	tools
forky_node_coverage	✅	14,938	5
tracoor_node_coverage	✅	16,591	6
mainnet_block_arrival_p50	✅	16,156	8
list_datasources	✅	13,025	2
block_count_24h	✅	16,599	9
missed_slots_24h	✅	20,045	16
chartkit_default_arrival_distribution	✅	36,197	12
storage_upload_session_scoped	✅	22,432	22

🔭 Langfuse traces (8 runs; ⚠️ = failed)

• forky_node_coverage
• tracoor_node_coverage
• mainnet_block_arrival_p50
• list_datasources
• block_count_24h
• missed_slots_24h
• chartkit_default_arrival_distribution
• storage_upload_session_scoped

_{The report walks this branch's commits against the master baseline and the most recent release. A self-contained copy is in the run's eval-smoke-* artifact.}