spec-decode for BitNet: subNorm+f32-KV batched verify, ?spec + ?bench=spec

core/engine.js

@@ -80,7 +80,7 @@ export async function createEngine(modelEntry, loaded) {
     // head, the drafter proposes and the target batch-verifies — output is BYTE-IDENTICAL to greedy decode
     // (greedy verify), streamed via onCommit. Any incompatibility/throw falls straight through to the standard
     // loop below, so default Q (no drafter) and unsupported models are completely unaffected.
-    if (!err && _drafter && gpu.specDecode && gpu.setDrafter) {
+    if (!err && (_drafter || globalThis.__spec) && gpu.specDecode && gpu.setDrafter) {
       try {
         gpu.setDrafter(_drafter);
         const out = []; const t0 = _perf(); let ttft2 = 0, tokps2 = 0;