per-pass GPU trace ?bench=trace: name the non-weight overhead

core/engine.js

@@ -185,5 +185,6 @@ export async function createEngine(modelEntry, loaded) {
     memoKey, memoGet: (k) => memo.get(k), memoHas: (k) => memo.has(k), memoSet: (k, v) => memo.set(k, v),
     buildReceipt, verify: verifyIntegrity, reDerive,
     stats: () => gpu.timing, reset: () => { try { gpu.reset(); } catch {} }, destroy: () => { try { gpu.destroy(); } catch {} },
+    _gpu: gpu,   // raw engine handle (diagnostics: per-pass GPU trace via the step() path + window.__profile)
   };
 }