perf(mempool): O(k) reap with head-offset compaction

[daps94]

Summary

• Replaces O(N) full txIndex rebuild after every Reap() with an O(k) head-offset design that advances a pointer past consumed entries
• Amortized compaction only triggers when head >= 1024 or head*2 >= len(txs), keeping the common path allocation-free
• Collapses separate exists() + push() into single-lock pushLocked(), eliminating a TOCTOU race on duplicate detection

Motivation

The ReapList maintained a txIndex map[string]int for deduplication. After every Reap(), the entire index was rebuilt from scratch (slices.DeleteFunc + re-index loop), even though Reap() only consumes entries from the front. This is O(N) in the total list size, wasted work when the list is large.

Benchmark Results (benchstat, 5 runs, Apple M4 Pro)

                                      │   main    │  reaplist   │
                                      │  sec/op   │  sec/op  vs │
EVMMempoolIterator/EVM_0/Cosmos_50       96.7µ      73.1µ   -24%
EVMMempoolIterator/EVM_0/Cosmos_500       224µ       150µ   -33%
EVMMempoolIterator/EVM_0/Cosmos_1000      321µ       258µ   -20%
EVMMempoolIterator/EVM_0/Cosmos_2500     1212µ       694µ   -43%
EVMMempoolIterator/EVM_0/Cosmos_10000    6101µ      4894µ   -20%
EVMMempoolIterator/EVM_1000/Cosmos_1k    4415µ      2504µ   -43%
EVMMempoolIterator/EVM_5000/Cosmos_5k   16965µ     14757µ   -13%

Geomean latency:  -21%

All Cosmos-path results statistically significant (p ≤ 0.032). EVM-only paths show no change (expected — ReapList is Cosmos-side only). Memory and allocs unchanged (zero-allocation design).

Files Changed

File	Change
mempool/reap_list.go	Add head int field; rewrite Reap() to iterate from head; add maybeCompactLocked(); merge exists()+push() into pushLocked()

Test Plan

• go test -tags=test -race ./mempool/... — 74 tests pass with race detection
• 27 focused reap_list tests pass
• Benchmark shows 21% geomean improvement (no regression in EVM-only paths)

🤖 Generated with Claude Code