Skip to content

2026-06-10-round2-hotpath-requirements.md

Latest commit

 

History

History
57 lines (44 loc) · 3.45 KB

2026-06-10-round2-hotpath-requirements.md

File metadata and controls

57 lines (44 loc) · 3.45 KB
Raw

Round-2 hot-path optimization campaign — requirements

Date: 2026-06-10 Status: approved (plan-mode approval of Round 2 section in the session plan file)

Summary

Land all three round-2 candidate tracks — the MuHash3072 ILP multiply rewrite, the sync_pipeline bench-validity fix plus apply-side cache repopulation, and the mimalloc bench A/B wiring — gated on correctness proof rather than measured wins. Criterion benches still run serially afterward as a regression veto, and an adversarial-verification pass audits each diff before its atomic commit.

Problem Frame

Round 1 of the same-node /optimize campaign closed all five non-script criterion benches with honest dispositions (three exit-12, two deferred). Round 2 generated three worktree-isolated candidates for the deferred targets, all compiling and test/clippy-clean. The blocking question was disposition policy: the T1 cache-repopulation win is production-only and unmeasurable by existing sub-benches, and the T2 MuHash rewrite is consensus-critical.

Key Decision — commit policy (user, verbatim)

Asked how to disposition T1 Part B given no measurement path, the user answered:

"We don't measure first, we optimize all the efforts when possible."

Combined with the twice-repeated directive "Run all the possible hot-path optimizations", this sets the round-2 policy: correctness-gated optimization-first. Correctness proof (byte-identity differentials, unit tests, clippy/fmt) decides commits; benches run only as a regression veto — a null result no longer blocks, a p<0.05 regression on a track's gate benches does.

Requirements

  • R1 (T2 MuHash): lands if the differential suite (frozen reference multiply, ≥100k seeded inputs, adversarial carry patterns), cargo test -p bitcoin-rs-coinstats, clippy, and fmt all pass re-run in the main tree. Commit perf(coinstats), flagged consensus-critical for user review.
  • R2 (T1 sync): Part A (bench-validity fix: timed regions cover only production-path work) commits first as fix(node). Part B (apply-side expected_apply_cache repopulation) commits on its cache-coherence structural argument plus unit tests (miss→populate→hit, invalidation on failed apply / tip move, horizon cap) as perf(node).
  • R3 (T3 mimalloc): commits as bench infrastructure (bench-mimalloc feature, plain dev-dependency, default OFF, zero src/ changes) if both configurations build/test/clippy clean. The A/B delta is recorded as a bench-fidelity finding. Only track adding a dependency — flagged and accepted.
  • R4 (regression veto): serial measurement barrier on idle pinned cores, one bench at a time: T2 vs saved baseline → T3 A/B → T1-A fixed-bench baseline → T1-B vs that baseline. Any p<0.05 regression on a track's gate benches blocks that track's commit.
  • R5 (adversarial verification): a parallel verification pass audits every track before commit: differential-gate re-execution, T1-B cache-coherence attack review, per-track diff-vs-claim audit.
  • R6 (commits): atomic, conventional, one concern each, Op: body trailers, no agent identity trailers; then worktree cleanup and disposition-table update.

Scope Boundaries

  • The cross-node faster-than-Core/gocoin verdict stays corpus-blocked (owner: user).
  • The three settled exit-12 targets (utxo_commit, sync_apply_metrics, kvstore_backends) stay closed.
  • No new sync_pipeline sub-bench — not needed as a commit gate under the round-2 policy.