Add regression test for scrypt non-determinism (#28607)

[robobun]

Regression test for #28607. The actual fix is oven-sh/WebKit#179 — which needs to merge upstream before this bug is closed. See below for why the Bun-side change is just a test.

Repro

import { hashPassword, verifyPassword } from 'better-auth/crypto';
const password = 'TestPassword123!';
const hash = await hashPassword(password);
await verifyPassword({ hash, password }); // intermittently false on Bun 1.3.11 (Arch)

Root cause

Undefined behavior in WebKit's IntegralTypedArrayAdaptor::toNativeFromDouble() (Source/JavaScriptCore/runtime/TypedArrayAdaptors.h):

static Type toNativeFromDouble(double value)
{
    int32_t result = static_cast<int32_t>(value);          // UB when value > INT32_MAX
    if (static_cast<double>(result) != value)
        result = toInt32(value);
    return static_cast<Type>(result);
}

static_cast<int32_t>(double) is UB in C++ when value is outside [INT32_MIN, INT32_MAX]. Clang 22 legally optimizes the subsequent range check away, silently turning every out-of-range typed-array store into INT32_MIN. That corrupts @noble/hashes/scrypt — and therefore Better Auth's hashPassword — producing different output for the same input across runs.

Diagnosed upstream by @carlsmedstad (Arch Linux maintainer) in #28607.

Fix

The fix lives in oven-sh/WebKit#179, which always routes through toInt32():

• On x86_64 it uses a cvttsd2si inline-asm fast path with a portable fallback to toIntImpl<int32_t>() when truncation isn't exact — avoiding the UB entirely.
• On ARM it keeps the existing FJCVTZS code path (which was already correct).
• The dead UB branch in TypedArrayAdaptors.h is removed.

This PR does not bump WEBKIT_VERSION. The earlier attempt (preview-pr-179-3f10db79) broke every C++ build lane because the WebKit Preview Build workflow on that PR is still in action_required state (the PR is from a fork) — no prebuilt binary exists. A WebKit maintainer needs to click Approve and run on https://github.com/oven-sh/WebKit/actions/runs/23983158093 before the tag publishes. Once that lands and oven-sh/WebKit#179 merges to main, a regular WebKit bump will pick up the fix and this test will guard it.

Test

test/regression/issue/28607.test.ts exercises the JS-visible behavior so the regression is caught by CI the next time a clang version triggers the UB:

• Int32Array stores of values > 2**31 must match v | 0 (ECMA-262 ToInt32 evaluated via a different WebKit code path).
• Uint32Array stores of values > 2**32 must match v >>> 0.
• Int16Array stores of values > 2**15 must wrap mod 2^16 with sign adjustment.
• The same inputs must produce the same Int32Array outputs across 50 repeated runs.
• A Salsa20-style add + rotate + xor mixing pattern on a 16-lane Int32Array (the exact shape that @noble/hashes/scrypt uses) must be deterministic across 50 runs.

Local bun bd test test/regression/issue/28607.test.ts passes all 5 tests / 54 expects.

Staying draft until WebKit PR #179 is merged so a version bump can pick it up.

[robobun]

^{Updated 1:44 AM PT - Apr 6th, 2026}

❌ @robobun, your commit 6a10324 has 2 failures in Build #43914 (All Failures):

• test/js/bun/webview/webview-chrome.test.ts - code 1 on 🍎 14 aarch64
• test/js/bun/util/v8-heap-snapshot.test.ts - SIGKILL on 🐧 25.04 x64-baseline

---

🧪   To try this PR locally:

bunx bun-pr 28905

That installs a local version of the PR into your bun-28905 executable, so you can run:

bun-28905 --bun