feat(capability): 转码能力包（vendor 纯 Go AMR-NB 解码器 + 黄金向量逐样本对照 + 8 哨兵 + 注册面）

internal/capability/transcode/amrnb/simd.go

@@ -0,0 +1,37 @@
+package amrnb
+
+// SIMD-accelerated non-saturating int16 FIR, used by the hot filter kernels.
+// The reference accumulates products into a plain int32 (no per-term
+// saturation), and integer addition is associative modulo 2^32, so accumulating
+// in the lane-parallel order an AVX2 kernel uses yields a bit-identical result.
+// firRawGeneric is the portable fallback and the oracle the assembly is
+// fuzz-tested against.
+
+// firRawGeneric computes the raw (pre-round, pre-shift) FIR sums
+// dst[o] = sum_t x[o+t]*coef[t] for every output o, accumulating in wrapping
+// int32. The amd64 build vectorises this across 8 outputs at a time; this is
+// the portable fallback and the fuzz-test oracle. Requires
+// len(x) >= len(dst)+len(coef)-1.
+func firRawGeneric(dst []int32, x []int16, coef []int16) {
+	taps := len(coef)
+	for o := range dst {
+		xw := x[o : o+taps]
+		var s int32
+		for t, c := range coef {
+			s += int32(xw[t]) * int32(c)
+		}
+		dst[o] = s
+	}
+}
+
+// firDotGeneric computes the wrapping-int32 dot product sum_i a[i]*b[i] over the
+// first len(a) elements (requires len(b) >= len(a)). The amd64 build vectorises
+// it with VPMADDWD; this is the portable fallback and the fuzz-test oracle.
+func firDotGeneric(a, b []int16) int32 {
+	bb := b[:len(a)]
+	var s int32
+	for i, av := range a {
+		s += int32(av) * int32(bb[i])
+	}
+	return s
+}