Parity benchmark: reproducible wasm→synth-AOT vs native-C→LLVM comparison — counter the 'wasm is ~40% slower' claim with measured cycles+bytes

[avrabe]

Motivation

A recurring industry claim: "WebAssembly is ~40% slower than native." That number comes from runtime/interpreter deployments (wasm3, interpreted sandboxes, JIT warm-up). It does not describe synth's model: AOT wasm→bare-metal native with no runtime — the output is Thumb-2/RV32 machine code, and the only question is compilation quality, which is measurable.

We have partial evidence both directions and should publish one honest, reproducible artifact instead of counter-slogans.

What we can already show (in-tree, CI-gated)

• Semantics: bit-exact vs wasmtime across the differential suite (incl. the full f32 set, 156/156 + 84/84).
• Structural wins native can't match: proof-carrying specialization (Beyond parity: proof-carrying specialization + optimal-alloc to BEAT LLVM (0.7× target) #494, SYNTH_FACT_SPEC) consumes loom wsc.facts value-range premises and ordeal-certifies guard/clamp elisions — gust_mix 50 B → 14 B on the branchless-select clamp, approaching LLVM's 12 B floor from a proven direction (LLVM cannot elide what it cannot prove; we carry the proof).
• Allocator at optimum on key kernels: flat_flight at its Belady frame-traffic-0 optimum since v0.24.0.

What we must measure, not assert

• Cycles on real silicon (i.MX RT1062 / STM32H743 class + qemu -icount cross-check): same kernel as (a) native C → arm-none-eabi-gcc/clang -Os and (b) wasm → synth compile (default + SYNTH_FACT_SPEC). The Beyond parity: proof-carrying specialization + optimal-alloc to BEAT LLVM (0.7× target) #494 kill-criterion (<1.0× then ≤0.70×) is gale's DWT measurement — OPEN.
• Size: the Codegen: close the 3.9× size gap vs LLVM on the gust hot path (regalloc + addressing-mode folding) #390 attribution shows synth still 3.56× on gust_poll (spill_reload ~31%) — the benchmark must show this honestly alongside the wins.

Deliverable

scripts/repro/parity_benchmark/ + artifacts/parity-benchmark.md:

1. Pinned toolchains, N kernels (gust_mix, gust_poll, flat_flight, a falcon f32 kernel), both compile paths, cycle+byte table.
2. Falsification instructions (one command per row).
3. Honest headline: where AOT-wasm ≥ native today (certified-elision shapes), where it isn't yet (general regalloc), and the trend per release.
4. CI-pinned so the numbers can't drift from prose (claims.yaml).

Refs: #494 (beat-LLVM), #390 (size attribution), #209 (perf epic), #242 (North Star). gale owns the silicon rows.

[avrabe]

Shipped in v0.42.0 (PR #737): scripts/repro/parity_benchmark/ + artifacts/parity-benchmark.md. Headline (conditioned): gust_mix clamp under a loom-proven premise = 14 B vs 26 B measured arm-none-eabi-gcc -Os (0.54× of native); unspecialized 84 B; LLVM floor cited; gaps shown (gust_poll 3.48×, flat_flight 2.54×, falcon-f32 1.64×); cycles all OPEN — gale silicon (DWT) owns those rows. One falsification command per row. Closing; cycle rows land under #494/#209 when measured. (— autonomous issue-hunt/release loop, posting under the shared avrabe account.)