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_amd64.go

@@ -0,0 +1,46 @@
+//go:build amd64
+
+package amrnb
+
+import "golang.org/x/sys/cpu"
+
+var useAVX2 = cpu.X86.HasAVX2
+
+// firRaw computes dst[o] = sum_t x[o+t]*coef[t] (wrapping int32). On AVX2 it
+// vectorises across 8 outputs per iteration; otherwise it falls back.
+func firRaw(dst []int32, x []int16, coef []int16) {
+	if useAVX2 && len(dst) >= 8 && len(coef) <= maxFIRTaps {
+		var c32 [maxFIRTaps]int32
+		for i, c := range coef {
+			c32[i] = int32(c)
+		}
+		firRawAVX2(dst, x, c32[:len(coef)])
+		return
+	}
+	firRawGeneric(dst, x, coef)
+}
+
+// maxFIRTaps bounds the stack-resident sign-extended coefficient buffer. It
+// covers the longest FIR used: the open-loop pitch correlation window, which
+// spans a whole frame (L_FRAME = 160 taps).
+const maxFIRTaps = 256
+
+// firRawAVX2 is implemented in simd_amd64.s. coef32 holds sign-extended taps.
+//
+//go:noescape
+func firRawAVX2(dst []int32, x []int16, coef32 []int32)
+
+// firDot computes the wrapping-int32 dot product sum_i a[i]*b[i] over len(a)
+// elements. On AVX2 it uses VPMADDWD across 16 int16 per iteration; otherwise it
+// falls back.
+func firDot(a, b []int16) int32 {
+	if useAVX2 && len(a) >= 16 {
+		return firDotAVX2(a, b)
+	}
+	return firDotGeneric(a, b)
+}
+
+// firDotAVX2 is implemented in simd_amd64.s. It reads len(a) int16 from a and b.
+//
+//go:noescape
+func firDotAVX2(a, b []int16) int32