feat(bb.js): paired A/B Chonk bench mode in browser-test-app

[AztecBot]

Adds a paired A/B benchmarking mode to the existing bb.js browser-test-app so PR-vs-base Chonk prove comparisons on real BrowserStack hardware produce statistically defensible numbers, reusing the harness we already trust instead of a separate minimal rig.

What it does

window.runChonkAb(options) (auto-runs on a base64url ?bench= param) runs a Chonk prove repeatedly, interleaving two bb.wasm variants on the same device in one session (alternating order per pair to remove first-mover bias), then reports the paired median Δ with a bootstrap 95% CI. Variant selection is at runtime via bb.js's existing wasmPath option — both binaries are hosted side by side, no rebuild and no second bundle:

bench-wasm/{base,pr}/barretenberg{,-threads}.wasm.gz

memMaxPages maps to Barretenberg.new({ memory: { maximum } }) (set 16384 = 1 GiB on iOS Safari, which rejects the default 4 GiB SAB).

Why

Between-session variance on real devices (different physical unit, thermal state, co-tenant) dwarfs single-digit prover deltas. Pairing both variants back-to-back on one device, N≥10 times, and reporting a CI is what distinguishes a real 3–6% delta from noise — or honestly reports that it cannot be distinguished. See BENCH.md.

Files

• src/index.ts — wasmPath/memory passthrough on proveChonk; new installChonkAbGlobal() (interleave loop, bootstrap-CI stats, /progress + /result POST, window.__benchResult).
• scripts/serve-bench.mjs — node server hosting dest/, /wasm/<variant>/, /inputs/<flow>/ivc-inputs.msgpack, /progress + /result sinks, with COOP/COEP for SAB. bb.js fetches CRS from the public CDN, so no CRS proxy.
• package.json — serve-bench script.
• BENCH.md — usage, the reporting rule (no signed delta when the CI crosses zero), and the iOS two-session workaround.

Verification status — DRAFT, not yet validated

• TypeScript transpiles cleanly (esbuild). Not yet full-built/type-checked against the bb.js portal types, and not yet BrowserStack smoke-run on this branch.
• Before this is marked ready: a full yarn build of the app + a paired-A/B smoke run on at least one desktop and one mobile BS target, confirming verified proofs and a populated paired block. I can run that as the next step.

Methodology background: https://gist.github.com/AztecBot/0483ec5b0293ff9e31621ad035854eb0 · protocol: https://gist.github.com/AztecBot/a8d3e5c14475c8fb5af520706e408222

This consolidates browser benching onto bb.js and supersedes the unmerged codex/browserstack-wasm-bench minimal harness (and barretenberg-claude#3799).

---

Created by claudebox · group: slackbot