@svrnsec/pulse 0.3.1 → 0.4.0

package/src/collector/entropy.js

@@ -0,0 +1,195 @@
+/**
+ * @sovereign/pulse — Entropy Collector
+ *
+ * Bridges the Rust/WASM matrix-multiply probe into JavaScript.
+ * The WASM module is lazily initialised once and cached for subsequent calls.
+ */
+
+import { collectEntropyAdaptive } from './adaptive.js';
+
+// ---------------------------------------------------------------------------
+// WASM loader (lazy singleton)
+// ---------------------------------------------------------------------------
+let _wasmModule   = null;
+let _initPromise  = null;
+
+/**
+ * Initialise (or return the cached) WASM module.
+ * Works in browsers (via fetch), in Electron (Node.js context), and in
+ * Jest/Vitest via a manual WASM path override.
+ *
+ * @param {string} [wasmPath] – override path/URL to the .wasm binary
+ */
+async function initWasm(wasmPath) {
+  if (_wasmModule) return _wasmModule;
+  if (_initPromise) return _initPromise;
+
+  _initPromise = (async () => {
+    // Dynamic import so bundlers can tree-shake this for server-only builds.
+    const { default: init, run_entropy_probe, run_memory_probe, compute_autocorrelation } =
+      await import('../../pkg/pulse_core.js');
+
+    const url = wasmPath ?? new URL('../../pkg/pulse_core_bg.wasm', import.meta.url).href;
+    await init(url);
+
+    _wasmModule = { run_entropy_probe, run_memory_probe, compute_autocorrelation };
+    return _wasmModule;
+  })();
+
+  return _initPromise;
+}
+
+// ---------------------------------------------------------------------------
+// collectEntropy
+// ---------------------------------------------------------------------------
+
+/**
+ * Run the WASM entropy probe and return raw timing data.
+ *
+ * @param {object}  opts
+ * @param {number}  [opts.iterations=200]  - number of matrix-multiply rounds
+ * @param {number}  [opts.matrixSize=64]   - N for the N×N matrices
+ * @param {number}  [opts.memSizeKb=512]   - size of the memory bandwidth probe
+ * @param {number}  [opts.memIterations=50]
+ * @param {boolean} [opts.phased=true]     - run cold/load/hot phases for entropy-jitter ratio
+ * @param {string}  [opts.wasmPath]        - optional custom WASM binary path
+ *
+ * @returns {Promise<EntropyResult>}
+ */
+export async function collectEntropy(opts = {}) {
+  const {
+    iterations        = 200,
+    matrixSize        = 64,
+    memSizeKb         = 512,
+    memIterations     = 50,
+    phased            = true,
+    adaptive          = false,
+    adaptiveThreshold = 0.85,
+    onBatch,
+    wasmPath,
+  } = opts;
+
+  const wasm    = await initWasm(wasmPath);
+  const t_start = Date.now();
+
+  let phases = null;
+  let timings, resolutionProbe, checksum, timerGranularityMs;
+  let _adaptiveInfo = null;
+
+  // ── Adaptive mode: smart early exit, fastest for obvious VMs ──────────
+  if (adaptive) {
+    const r = await collectEntropyAdaptive(wasm, {
+      minIterations:   50,
+      maxIterations:   iterations,
+      batchSize:       25,
+      vmThreshold:     adaptiveThreshold,
+      hwThreshold:     0.80,
+      hwMinIterations: 75,
+      matrixSize,
+      onBatch,
+    });
+    timings            = r.timings;
+    resolutionProbe    = r.resolutionProbe ?? [];
+    checksum           = r.checksum;
+    timerGranularityMs = r.timerGranularityMs;
+    _adaptiveInfo      = { earlyExit: r.earlyExit, batches: r.batches, elapsedMs: r.elapsedMs };
+
+  // ── Phased collection: cold → load → hot ──────────────────────────────
+  // Each phase runs a separate WASM probe.  On real hardware, sustained load
+  // increases thermal noise so Phase 3 (hot) entropy is measurably higher
+  // than Phase 1 (cold).  A VM's hypervisor clock is insensitive to guest
+  // thermal state, so all three phases return nearly identical entropy.
+  } else if (phased && iterations >= 60) {
+    const coldN = Math.floor(iterations * 0.25);  // ~25% cold
+    const loadN = Math.floor(iterations * 0.50);  // ~50% sustained load
+    const hotN  = iterations - coldN - loadN;     // ~25% hot
+
+    const cold = wasm.run_entropy_probe(coldN, matrixSize);
+    const load = wasm.run_entropy_probe(loadN, matrixSize);
+    const hot  = wasm.run_entropy_probe(hotN,  matrixSize);
+
+    const coldTimings = Array.from(cold.timings);
+    const loadTimings = Array.from(load.timings);
+    const hotTimings  = Array.from(hot.timings);
+
+    timings         = [...coldTimings, ...loadTimings, ...hotTimings];
+    resolutionProbe = Array.from(cold.resolution_probe);
+    checksum        = (cold.checksum + load.checksum + hot.checksum).toString();
+
+    const { detectQuantizationEntropy } = await import('../analysis/jitter.js');
+    const coldQE = detectQuantizationEntropy(coldTimings);
+    const hotQE  = detectQuantizationEntropy(hotTimings);
+
+    phases = {
+      cold: { n: coldN, timings: coldTimings, qe: coldQE, mean: _mean(coldTimings) },
+      load: { n: loadN, timings: loadTimings, qe: detectQuantizationEntropy(loadTimings), mean: _mean(loadTimings) },
+      hot:  { n: hotN,  timings: hotTimings,  qe: hotQE,  mean: _mean(hotTimings)  },
+      // The key signal: entropy growth under load.
+      // Real silicon: hotQE / coldQE typically 1.05 – 1.40
+      // VM:           hotQE / coldQE typically 0.95 – 1.05 (flat)
+      entropyJitterRatio: coldQE > 0 ? hotQE / coldQE : 1.0,
+    };
+  } else {
+    // Single-phase fallback (fewer iterations or phased disabled)
+    const result    = wasm.run_entropy_probe(iterations, matrixSize);
+    timings         = Array.from(result.timings);
+    resolutionProbe = Array.from(result.resolution_probe);
+    checksum        = result.checksum.toString();
+  }
+
+  // ── Timer resolution (non-adaptive path only — adaptive computes its own) ─
+  if (!adaptive) {
+    const resDeltas = [];
+    for (let i = 1; i < resolutionProbe.length; i++) {
+      const d = resolutionProbe[i] - resolutionProbe[i - 1];
+      if (d > 0) resDeltas.push(d);
+    }
+    timerGranularityMs = resDeltas.length
+      ? resDeltas.reduce((a, b) => Math.min(a, b), Infinity)
+      : null;
+  }
+
+  // ── Autocorrelation at diagnostic lags ────────────────────────────────
+  // Extended lags catch long-period steal-time rhythms (Xen: ~150 iters)
+  const lags = [1, 2, 3, 5, 10, 25, 50];
+  const autocorrelations = {};
+  for (const lag of lags) {
+    if (lag < timings.length) {
+      autocorrelations[`lag${lag}`] = wasm.compute_autocorrelation(timings, lag);
+    }
+  }
+
+  // ── Secondary probe: memory bandwidth jitter ───────────────────────────
+  const memTimings = Array.from(wasm.run_memory_probe(memSizeKb, memIterations));
+
+  return {
+    timings,
+    resolutionProbe,
+    timerGranularityMs,
+    autocorrelations,
+    memTimings,
+    phases,
+    checksum,
+    collectedAt: t_start,
+    iterations: timings.length,   // actual count (adaptive may differ from requested)
+    matrixSize,
+    adaptive: _adaptiveInfo,      // null in non-adaptive mode
+  };
+}
+
+function _mean(arr) {
+  return arr.length ? arr.reduce((s, v) => s + v, 0) / arr.length : 0;
+}
+
+/**
+ * @typedef {object} EntropyResult
+ * @property {number[]}  timings            - per-iteration wall-clock deltas (ms)
+ * @property {number[]}  resolutionProbe    - raw successive perf.now() readings
+ * @property {number|null} timerGranularityMs - effective timer resolution
+ * @property {object}    autocorrelations   - { lag1, lag2, lag3, lag5, lag10 }
+ * @property {number[]}  memTimings         - memory-probe timings (ms)
+ * @property {string}    checksum           - proof the computation ran
+ * @property {number}    collectedAt        - Date.now() at probe start
+ * @property {number}    iterations
+ * @property {number}    matrixSize
+ */

package/src/collector/gpu.js

@@ -0,0 +1,245 @@
+/**
+ * @sovereign/pulse — WebGPU Thermal Variance Probe
+ *
+ * Runs a compute shader on the GPU and measures dispatch timing variance.
+ *
+ * Why this works
+ * ──────────────
+ * Real consumer GPUs (GTX 1650, RX 6600, M2 GPU) have thermal noise in shader
+ * execution timing that increases under sustained load — the same thermodynamic
+ * principle as the CPU probe but in silicon designed for parallel throughput.
+ *
+ * Cloud VMs with software GPU emulation (SwiftShader, llvmpipe, Mesa's softpipe)
+ * execute shaders on the CPU and produce near-deterministic timing — flat CV,
+ * no thermal growth across phases, no dispatch jitter.
+ *
+ * VMs with GPU passthrough (rare in practice, requires dedicated hardware) pass
+ * this check — which is correct, they have real GPU silicon.
+ *
+ * Signals
+ * ───────
+ *   gpuPresent      false = WebGPU absent = software renderer = high VM probability
+ *   isSoftware      true  = SwiftShader/llvmpipe detected by adapter info
+ *   dispatchCV      coefficient of variation across dispatch timings
+ *   thermalGrowth   (hotDispatchMean - coldDispatchMean) / coldDispatchMean
+ *   vendorString    GPU vendor from adapter info (Intel, NVIDIA, AMD, Apple, etc.)
+ */
+
+/* ─── WebGPU availability ────────────────────────────────────────────────── */
+
+function isWebGPUAvailable() {
+  return typeof navigator !== 'undefined' && 'gpu' in navigator;
+}
+
+/* ─── Software renderer detection ───────────────────────────────────────── */
+
+const SOFTWARE_RENDERER_PATTERNS = [
+  /swiftshader/i,
+  /llvmpipe/i,
+  /softpipe/i,
+  /microsoft basic render/i,
+  /angle \(.*software/i,
+  /cpu/i,
+];
+
+function detectSoftwareRenderer(adapterInfo) {
+  const desc = [
+    adapterInfo?.vendor   ?? '',
+    adapterInfo?.device   ?? '',
+    adapterInfo?.description ?? '',
+    adapterInfo?.architecture ?? '',
+  ].join(' ');
+
+  return SOFTWARE_RENDERER_PATTERNS.some(p => p.test(desc));
+}
+
+/* ─── Compute shader ─────────────────────────────────────────────────────── */
+
+// A compute workload that is trivially parallelisable but forces the GPU to
+// actually execute — matrix-multiply on 64 × 64 tiles across 256 workgroups.
+// Light enough that it doesn't block UI; heavy enough to generate thermal signal.
+const SHADER_SRC = /* wgsl */ `
+  struct Matrix {
+    values: array<f32, 4096>,  // 64x64
+  };
+
+  @group(0) @binding(0) var<storage, read>       matA : Matrix;
+  @group(0) @binding(1) var<storage, read>       matB : Matrix;
+  @group(0) @binding(2) var<storage, read_write> matC : Matrix;
+
+  @compute @workgroup_size(8, 8)
+  fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let row = gid.x;
+    let col = gid.y;
+    if (row >= 64u || col >= 64u) { return; }
+
+    var acc: f32 = 0.0;
+    for (var k = 0u; k < 64u; k++) {
+      acc += matA.values[row * 64u + k] * matB.values[k * 64u + col];
+    }
+    matC.values[row * 64u + col] = acc;
+  }
+`;
+
+/* ─── collectGpuEntropy ─────────────────────────────────────────────────── */
+
+/**
+ * @param {object}  [opts]
+ * @param {number}  [opts.iterations=60]      – dispatch rounds per phase
+ * @param {boolean} [opts.phased=true]         – cold / load / hot phases
+ * @param {number}  [opts.timeoutMs=8000]      – hard abort if GPU stalls
+ * @returns {Promise<GpuEntropyResult>}
+ */
+export async function collectGpuEntropy(opts = {}) {
+  const { iterations = 60, phased = true, timeoutMs = 8000 } = opts;
+
+  if (!isWebGPUAvailable()) {
+    return _noGpu('WebGPU not available in this environment');
+  }
+
+  let adapter, device;
+  try {
+    adapter = await Promise.race([
+      navigator.gpu.requestAdapter({ powerPreference: 'high-performance' }),
+      _timeout(timeoutMs, 'requestAdapter timed out'),
+    ]);
+    if (!adapter) return _noGpu('No WebGPU adapter found');
+
+    device = await Promise.race([
+      adapter.requestDevice(),
+      _timeout(timeoutMs, 'requestDevice timed out'),
+    ]);
+  } catch (err) {
+    return _noGpu(`WebGPU init failed: ${err.message}`);
+  }
+
+  const adapterInfo = adapter.info ?? {};
+  const isSoftware  = detectSoftwareRenderer(adapterInfo);
+
+  // Compile the shader module once
+  const shaderModule = device.createShaderModule({ code: SHADER_SRC });
+
+  // Create persistent GPU buffers (64×64 float32 = 16 KB each)
+  const bufSize = 4096 * 4; // 4096 floats × 4 bytes
+  const bufA = _createBuffer(device, bufSize, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST);
+  const bufB = _createBuffer(device, bufSize, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST);
+  const bufC = _createBuffer(device, bufSize, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC);
+
+  // Seed with random data
+  const matData = new Float32Array(4096).map(() => Math.random());
+  device.queue.writeBuffer(bufA, 0, matData);
+  device.queue.writeBuffer(bufB, 0, matData);
+
+  const pipeline = device.createComputePipeline({
+    layout: 'auto',
+    compute: { module: shaderModule, entryPoint: 'main' },
+  });
+
+  const bindGroup = device.createBindGroup({
+    layout: pipeline.getBindGroupLayout(0),
+    entries: [
+      { binding: 0, resource: { buffer: bufA } },
+      { binding: 1, resource: { buffer: bufB } },
+      { binding: 2, resource: { buffer: bufC } },
+    ],
+  });
+
+  // ── Probe ──────────────────────────────────────────────────────────────
+  async function runPhase(n) {
+    const timings = [];
+    for (let i = 0; i < n; i++) {
+      const t0 = performance.now();
+      const encoder = device.createCommandEncoder();
+      const pass    = encoder.beginComputePass();
+      pass.setPipeline(pipeline);
+      pass.setBindGroup(0, bindGroup);
+      pass.dispatchWorkgroups(8, 8); // 64 workgroups total
+      pass.end();
+      device.queue.submit([encoder.finish()]);
+      await device.queue.onSubmittedWorkDone();
+      const t1 = performance.now();
+      timings.push(t1 - t0);
+    }
+    return timings;
+  }
+
+  let coldTimings, loadTimings, hotTimings;
+
+  if (phased) {
+    coldTimings = await runPhase(Math.floor(iterations * 0.25));
+    loadTimings = await runPhase(Math.floor(iterations * 0.50));
+    hotTimings  = await runPhase(iterations - coldTimings.length - loadTimings.length);
+  } else {
+    coldTimings = await runPhase(iterations);
+    loadTimings = [];
+    hotTimings  = [];
+  }
+
+  // Cleanup
+  bufA.destroy(); bufB.destroy(); bufC.destroy();
+  device.destroy();
+
+  const allTimings = [...coldTimings, ...loadTimings, ...hotTimings];
+  const mean       = _mean(allTimings);
+  const cv         = mean > 0 ? _std(allTimings) / mean : 0;
+
+  const coldMean = _mean(coldTimings);
+  const hotMean  = _mean(hotTimings.length ? hotTimings : coldTimings);
+  const thermalGrowth = coldMean > 0 ? (hotMean - coldMean) / coldMean : 0;
+
+  return {
+    gpuPresent:    true,
+    isSoftware,
+    vendor:        adapterInfo.vendor      ?? 'unknown',
+    architecture:  adapterInfo.architecture ?? 'unknown',
+    timings:       allTimings,
+    dispatchCV:    cv,
+    thermalGrowth,
+    coldMean,
+    hotMean,
+    // Heuristic: real GPU → thermalGrowth > 0.02 and CV > 0.04
+    // Software renderer → thermalGrowth ≈ 0, CV < 0.02
+    verdict: isSoftware ? 'software_renderer'
+      : thermalGrowth > 0.02 && cv > 0.04 ? 'real_gpu'
+      : thermalGrowth < 0 && cv < 0.02   ? 'virtual_gpu'
+      : 'ambiguous',
+  };
+}
+
+/* ─── helpers ────────────────────────────────────────────────────────────── */
+
+function _noGpu(reason) {
+  return { gpuPresent: false, isSoftware: false, vendor: null,
+           architecture: null, timings: [], dispatchCV: 0,
+           thermalGrowth: 0, coldMean: 0, hotMean: 0,
+           verdict: 'no_gpu', reason };
+}
+
+function _createBuffer(device, size, usage) {
+  return device.createBuffer({ size, usage });
+}
+
+function _mean(arr) {
+  return arr.length ? arr.reduce((s, v) => s + v, 0) / arr.length : 0;
+}
+
+function _std(arr) {
+  const m = _mean(arr);
+  return Math.sqrt(arr.reduce((s, v) => s + (v - m) ** 2, 0) / arr.length);
+}
+
+function _timeout(ms, msg) {
+  return new Promise((_, reject) => setTimeout(() => reject(new Error(msg)), ms));
+}
+
+/**
+ * @typedef {object} GpuEntropyResult
+ * @property {boolean}  gpuPresent
+ * @property {boolean}  isSoftware
+ * @property {string|null} vendor
+ * @property {string|null} architecture
+ * @property {number[]}  timings
+ * @property {number}    dispatchCV
+ * @property {number}    thermalGrowth
+ * @property {string}    verdict   'real_gpu' | 'virtual_gpu' | 'software_renderer' | 'no_gpu' | 'ambiguous'
+ */

package/src/collector/sabTimer.js

@@ -0,0 +1,191 @@
+/**
+ * @sovereign/pulse — SharedArrayBuffer Microsecond Timer
+ *
+ * Bypasses browser timer clamping (Brave 100µs cap, Firefox 20µs cap, Safari
+ * 1ms cap) using Atomics.wait() which is exempt from clamping because it maps
+ * directly to OS-level futex/semaphore primitives.
+ *
+ * Requirements
+ * ────────────
+ * The page must be served with Cross-Origin Isolation headers:
+ *   Cross-Origin-Opener-Policy: same-origin
+ *   Cross-Origin-Embedder-Policy: require-corp
+ *
+ * These are mandatory for security (Spectre mitigations) and are already
+ * required by WebGPU, WebAssembly threads, and SharedArrayBuffer in all
+ * modern browsers.
+ *
+ * What we measure
+ * ───────────────
+ *   resolution        the true timer resolution (pre-clamp) in microseconds
+ *   isClamped         true if performance.now() is artificially reduced
+ *   clampAmount       how much performance.now() was rounded (µs)
+ *   highResTimings    entropy probe timings at true microsecond resolution
+ *
+ * Why this matters
+ * ────────────────
+ * With 1ms clamping, a VM's flat distribution and a real device's noisy
+ * distribution can look similar — both get quantized to the same step.
+ * At 1µs resolution, the difference between EJR=1.01 and EJR=1.24 is
+ * unmistakable. This upgrade alone materially improves detection accuracy
+ * on Brave and Firefox where timer clamping was previously a confound.
+ */
+
+/* ─── availability ───────────────────────────────────────────────────────── */
+
+export function isSabAvailable() {
+  return (
+    typeof SharedArrayBuffer !== 'undefined' &&
+    typeof Atomics            !== 'undefined' &&
+    typeof Atomics.wait       === 'function'  &&
+    crossOriginIsolated === true              // window flag set by COOP+COEP headers
+  );
+}
+
+/* ─── Atomics-based high-resolution clock ───────────────────────────────── */
+
+let _sab  = null;
+let _i32  = null;
+
+function _initSab() {
+  if (!_sab) {
+    _sab = new SharedArrayBuffer(4);
+    _i32 = new Int32Array(_sab);
+  }
+}
+
+/**
+ * Wait exactly `us` microseconds using Atomics.wait().
+ * Returns wall-clock elapsed in milliseconds.
+ * Much more accurate than setTimeout(fn, 0) or performance.now() loops.
+ *
+ * @param {number} us – microseconds to wait
+ * @returns {number}  actual elapsed ms
+ */
+function _atomicsWait(us) {
+  _initSab();
+  const t0 = performance.now();
+  Atomics.wait(_i32, 0, 0, us / 1000); // Atomics.wait timeout is in ms
+  return performance.now() - t0;
+}
+
+/* ─── measureClamp ───────────────────────────────────────────────────────── */
+
+/**
+ * Determine the true timer resolution by comparing a series of
+ * sub-millisecond Atomics.wait() calls against performance.now() deltas.
+ *
+ * @returns {{ isClamped: boolean, clampAmountUs: number, resolutionUs: number }}
+ */
+export function measureClamp() {
+  if (!isSabAvailable()) {
+    return { isClamped: false, clampAmountUs: 0, resolutionUs: 1000 };
+  }
+
+  // Measure the minimum non-zero performance.now() delta
+  const performanceDeltas = [];
+  for (let i = 0; i < 100; i++) {
+    const t0 = performance.now();
+    let t1 = t0;
+    while (t1 === t0) t1 = performance.now();
+    performanceDeltas.push((t1 - t0) * 1000); // convert to µs
+  }
+  performanceDeltas.sort((a, b) => a - b);
+  const perfResolutionUs = performanceDeltas[Math.floor(performanceDeltas.length * 0.1)]; // 10th percentile
+
+  // Measure actual OS timer resolution via Atomics.wait
+  const atomicsDeltas = [];
+  for (let i = 0; i < 20; i++) {
+    const elapsedMs = _atomicsWait(100); // wait 100µs
+    atomicsDeltas.push(Math.abs(elapsedMs * 1000 - 100)); // error from target
+  }
+  const atomicsErrorUs = atomicsDeltas.reduce((s, v) => s + v, 0) / atomicsDeltas.length;
+  const trueResolutionUs = Math.max(1, atomicsErrorUs);
+
+  const isClamped    = perfResolutionUs > trueResolutionUs * 5;
+  const clampAmountUs = isClamped ? perfResolutionUs - trueResolutionUs : 0;
+
+  return { isClamped, clampAmountUs, resolutionUs: perfResolutionUs };
+}
+
+/* ─── collectHighResTimings ──────────────────────────────────────────────── */
+
+/**
+ * Collect entropy probe timings at Atomics-level resolution.
+ * Falls back to performance.now() if SAB is unavailable.
+ *
+ * The probe itself is identical to the WASM matrix probe — CPU work unit
+ * timed with the highest available clock. The difference: on a clamped
+ * browser this replaces quantized 100µs buckets with true µs measurements.
+ *
+ * @param {object} opts
+ * @param {number} [opts.iterations=200]
+ * @param {number} [opts.matrixSize=32]     – smaller than WASM probe (no SIMD here)
+ * @returns {{ timings: number[], usingAtomics: boolean, resolutionUs: number }}
+ */
+export function collectHighResTimings(opts = {}) {
+  const { iterations = 200, matrixSize = 32 } = opts;
+
+  const usingAtomics = isSabAvailable();
+  const clampInfo    = usingAtomics ? measureClamp() : { resolutionUs: 1000 };
+
+  // Simple matrix multiply work unit (JS — no WASM needed for the clock probe)
+  const N = matrixSize;
+  const A = new Float64Array(N * N).map(() => Math.random());
+  const B = new Float64Array(N * N).map(() => Math.random());
+  const C = new Float64Array(N * N);
+
+  const timings = new Array(iterations);
+
+  for (let iter = 0; iter < iterations; iter++) {
+    C.fill(0);
+
+    if (usingAtomics) {
+      // ── Atomics path: start timing, do work, read Atomics-calibrated time ──
+      // We use a sliding window approach: measure with Atomics.wait(0) which
+      // returns immediately but the OS schedules give us a high-res timestamp
+      // via the before/after pattern on the shared memory notification.
+      _initSab();
+
+      const tAtomicsBefore = _getAtomicsTs();
+      for (let i = 0; i < N; i++) {
+        for (let k = 0; k < N; k++) {
+          const aik = A[i * N + k];
+          for (let j = 0; j < N; j++) C[i * N + j] += aik * B[k * N + j];
+        }
+      }
+      const tAtomicsAfter = _getAtomicsTs();
+      timings[iter] = (tAtomicsAfter - tAtomicsBefore) * 1000; // µs → ms
+
+    } else {
+      // ── Standard path: use performance.now() ──
+      const t0 = performance.now();
+      for (let i = 0; i < N; i++) {
+        for (let k = 0; k < N; k++) {
+          const aik = A[i * N + k];
+          for (let j = 0; j < N; j++) C[i * N + j] += aik * B[k * N + j];
+        }
+      }
+      timings[iter] = performance.now() - t0;
+    }
+  }
+
+  return {
+    timings,
+    usingAtomics,
+    resolutionUs: clampInfo.resolutionUs,
+    isClamped:    clampInfo.isClamped ?? false,
+    clampAmountUs: clampInfo.clampAmountUs ?? 0,
+  };
+}
+
+/* ─── internal Atomics timestamp ─────────────────────────────────────────── */
+
+// Use a write to shared memory + memory fence as a timestamp anchor.
+// This forces the CPU to flush its store buffer, giving a hardware-ordered
+// time reference that survives compiler reordering.
+function _getAtomicsTs() {
+  _initSab();
+  Atomics.store(_i32, 0, Atomics.load(_i32, 0) + 1);
+  return performance.now();
+}