execution: reduce interpreter, syscall and runtime overhead

[7layermagik]

Reduce CPU work and allocations during transaction execution: speed up the sBPF interpreter, remove SHA-256 and memory-syscall overhead, and accelerate LtHash account updates. Exact timing remains unchanged.

This PR targets 7layer/review-performance-core so the earlier validator-performance changes stay outside its diff. Retarget to alpenglow-dev after those prerequisites merge.

Implementation

• Interpreter: keep hot operations together, use fixed register storage, pre-resolve internal calls, and reduce metering and memory-translation overhead. Account for the call-target table in program-cache memory usage.
• SHA-256: decode slice descriptors directly and write the digest without temporary allocations. Include the VM bounds-overflow correction required by differential testing.
• Memory syscalls: remove temporary copy buffers and accelerate comparison/fill operations. Preserve CU charges, error precedence, zero-length behavior, memcpy overlap rejection, memmove overlap support, and copy-on-write/growing regions.
• LtHash: use runtime-gated AVX2 with portable and purego fallbacks, preserving 16-bit wrapping and aliased operands.
• Fix the VASA test call signatures and Noop fixture's compute meter. Retain reusable differential tests and benchmarks; remove the abandoned SHA buffering prototype and investigation narration.

Claude's original commits retain their authorship. ARSH/verifier semantics changes, incremental execution and account prefetch are outside this PR. Timing sampling is deferred because its implementation introduced measurable overhead even in exact mode.

Measurements

Zen 5 / Ryzen 7 9700X, Go 1.26.4, GOMAXPROCS=1, CPU affinity and alternating-order repeated runs on a shared validator host:

Component	Before → after
Interpreter: Token-2022 transfer	19.52 → 13.70 µs (1.42×)
Interpreter: arithmetic, 5,000 iterations	167.24 → 94.55 µs (1.77×)
Interpreter: CPI to System Allocate	16.75 → 13.77 µs (1.22×)
SHA syscall in captured hash loop	234.5 → 165.9 ns/hash
Memory syscalls: Token-2022 allocations	4 fewer allocations / 192 fewer bytes per invocation
LtHash MixIn: portable → AVX2	approximately 310 → 15 ns, zero allocations

These are separate component comparisons, not cumulative speedups. Interpreter measurements compare 9db7dcb plus the same harness against its interpreter-patched counterpart 2ed5e533; SHA compares the frozen original syscall with the optimized implementation. Memory/LtHash measurements use staged commits e791b7bb and e4100766.

Whole replay did not establish an additional latency improvement. The later live baseline, f6e3fab1, already includes the interpreter and SHA optimizations. Replaying the same 233 blocks took 34.256 and 33.872 seconds for candidate 60d8e96b; four baseline runs ranged 33.378–34.273 seconds. Candidate mean was 0.82% higher. Shared-host variation and single-core tests do not establish improved multicore contention or voting speed.

Three post-deployment five-minute windows had combined FAST inclusion of 96.00% versus 96.70% before; observed completion-to-local-vote-enqueue p99 was 105.9 versus 111.9 ms. Leaders, workloads and host activity were unmatched, so neither difference establishes causation.

Correctness and reproduction

• 100,000 generated programs match reference return values, errors, CU usage and memory; pool-reuse checks pass.
• All 233 replayed Alpenglow blocks have identical normalized slot sets and bank hashes.
• Affected unit suites, targeted race tests and vet pass. LtHash randomized, wrapping, aliased, unaligned, runtime-fallback and purego checks pass, as do ARM64 compilation and native ARM64 LtHash tests.
• Legacy BPF-loader fixture failures reproduce on both baseline and candidate; this is not a full sealevel-suite pass.
• Final cleanup changes only documentation/test code. Targeted SHA/program tests, SHA race tests and benchmark smoke checks pass.

See execution benchmark instructions and SHA benchmark instructions for commands, pinned fixtures and measurement boundaries. Raw run outputs remain outside the merge diff.

The earlier candidate measurements above predate this regrouping. The new base aggregates the separately reviewed prerequisite branches; it is a testing/integration branch, not another merge PR. This PR now also owns boolean CLI/TOML default resolution, retained vote-deque ownership, SBPF v2 opcode/differential coverage, the pooled-heap dirty-bitmap bound, zero-length memcpy handling, and sibling-header write tracking. Claude's authored commits remain intact.

Current regrouping validation: affected SBPF/LtHash tests and focused memory/SHA/sibling/deque/config tests pass, as do build and vet. Broader sealevel fixture selection still hits legacy missing-compute-meter/setup failures; no full sealevel pass is claimed. No new Zen 5 measurement or deployment was performed for this regrouping. The former PR head is preserved on 7layer/archive-execution-before-regroup-20260924.

Review dependencies

One of four grouped reviews: Turbine/production #287 and voting #288 form the integration base for this PR; replay/recovery/streaming #292 builds on this PR. The integration base is not an additional PR. All 22 execution commits are unchanged by range-diff; the rebased execution binary builds.