fix(verify): enlarge fuzz buffers so slice-based kernels can be fuzzed

directly

verify/fuzz.go

@@ -221,7 +221,10 @@ func genDualArgs(rng *rand.Rand, sig funcSig, argSize int) (gasmArgs, goArgs []b
 			} else {
 				declaredLen = n + 512
 			}
-			bufBytes := (declaredLen+16)*elemSize + 128
+			// Allocate a buffer comfortably larger than declaredLen*elemSize so
+			// that SIMD over-reads and functions that write slightly past len
+			// (e.g. decoders that trust len(src)) never touch unmapped memory.
+			bufBytes := declaredLen*elemSize + 8192
 			// Two independent buffers with identical random content.
 			buf1 := make([]byte, bufBytes)
 			buf2 := make([]byte, bufBytes)
@@ -230,6 +233,8 @@ func genDualArgs(rng *rand.Rand, sig funcSig, argSize int) (gasmArgs, goArgs []b
 			bufs = append(bufs, buf1, buf2)
 			putPtr(gasmArgs, off, unsafe.Pointer(&buf1[0]))
 			putPtr(goArgs, off, unsafe.Pointer(&buf2[0]))
+			// len and cap both equal declaredLen — the buffer is guaranteed
+			// to hold at least declaredLen elements plus safety margin.
 			putU64(gasmArgs, off+8, uint64(declaredLen))
 			putU64(gasmArgs, off+16, uint64(declaredLen))
 			putU64(goArgs, off+8, uint64(declaredLen))