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/dsp_pstfilt.go

@@ -0,0 +1,108 @@
+package amrnb
+
+// Adaptive formant post-filter, ported from opencore-amrnb pstfilt.cpp. For each
+// subframe it builds the weighted LP filters A(z/γ3) and A(z/γ4), forms the
+// residual through A(z/γ3), derives a spectral-tilt compensation coefficient
+// from the truncated impulse response, applies it via preemphasis, synthesizes
+// through 1/A(z/γ4), and finally matches the energy back to the synthesis with
+// AGC. Uses the building blocks weightAi/residu/synFilt/preemphasis/agc; this is
+// the orchestration that chains them. Bit-exact with the reference.
+
+const (
+	cL_H = 22    // truncated impulse-response length
+	cMU  = 26214 // tilt compensation factor (0.8)
+)
+
+// γ bandwidth-expansion factors (pstfilt.cpp).
+var gamma3MR122 = [cM]int16{22938, 16057, 11240, 7868, 5508, 3856, 2699, 1889, 1322, 925}
+var gamma3 = [cM]int16{18022, 9912, 5451, 2998, 1649, 907, 499, 274, 151, 83}
+var gamma4MR122 = [cM]int16{24576, 18432, 13824, 10368, 7776, 5832, 4374, 3281, 2461, 1846}
+var gamma4 = [cM]int16{22938, 16057, 11240, 7868, 5508, 3856, 2699, 1889, 1322, 925}
+
+// postFilterState holds the formant post-filter memory across frames.
+type postFilterState struct {
+	res2      [cL_SUBFR]int16
+	memSynPst [cM]int16
+	preemph   int16
+	agc       agcState
+	synthBuf  [cM + cL_FRAME]int16 // M history + current frame
+}
+
+func (st *postFilterState) reset() {
+	*st = postFilterState{}
+}
+
+// postFilter post-filters one 160-sample synthesis frame in place. az4 holds the
+// four per-subframe LP coefficient sets (4*cMP1). Mirrors Post_Filter.
+func (st *postFilterState) postFilter(mode int, syn []int16, az4 []int16) {
+	var ap3, ap4 [cMP1]int16
+	var h [cL_H]int16
+
+	synWork := st.synthBuf[cM:] // syn_work[i] == synthBuf[M+i]
+	copy(synWork[:cL_FRAME], syn[:cL_FRAME])
+
+	azOff := 0
+	for iSubfr := 0; iSubfr < cL_FRAME; iSubfr += cL_SUBFR {
+		az := az4[azOff:]
+		if mode == dMR122 || mode == dMR102 {
+			weightAi(az, gamma3MR122[:], ap3[:])
+			weightAi(az, gamma4MR122[:], ap4[:])
+		} else {
+			weightAi(az, gamma3[:], ap3[:])
+			weightAi(az, gamma4[:], ap4[:])
+		}
+
+		// Residual through A(z/γ3); syn_work[iSubfr] == synthBuf[M+iSubfr].
+		residu(ap3[:], st.synthBuf[:], cM+iSubfr, cL_SUBFR, st.res2[:])
+
+		// Truncated impulse response of A(z/γ3)/A(z/γ4).
+		copy(h[:cMP1], ap3[:cMP1])
+		for i := cMP1; i < cL_H; i++ {
+			h[i] = 0
+		}
+		synFilt(ap4[:], h[:], h[:], cL_H, h[cMP1:], false)
+
+		// Tilt compensation coefficient from the impulse-response autocorr.
+		var lTmp int32
+		for i := cL_H - 1; i >= 0; i-- {
+			t2 := int32(h[i]) * int32(h[i])
+			if t2 == 0x40000000 {
+				break // reference sets overflow and stops without adding
+			}
+			lTmp = L_add(lTmp, t2<<1)
+		}
+		temp1 := int16(lTmp >> 16)
+
+		lTmp = 0
+		for i := cL_H - 2; i >= 0; i-- {
+			t2 := int32(h[i]) * int32(h[i+1])
+			if t2 == 0x40000000 {
+				break
+			}
+			lTmp = L_add(lTmp, t2<<1)
+		}
+		temp2 := int16(lTmp >> 16)
+
+		if temp2 <= 0 {
+			temp2 = 0
+		} else {
+			lt := (int32(temp2) * cMU) >> 15
+			if lt&0x00010000 != 0 {
+				lt |= int32(-0x10000)
+			}
+			temp2 = int16(lt)
+			temp2 = div_s(temp2, temp1)
+		}
+
+		preemphasis(&st.preemph, st.res2[:], temp2, cL_SUBFR)
+
+		synFilt(ap4[:], st.res2[:], syn[iSubfr:], cL_SUBFR, st.memSynPst[:], true)
+
+		st.agc.agc(synWork[iSubfr:], syn[iSubfr:], cAGC_FAC, cL_SUBFR)
+
+		azOff += cMP1
+	}
+
+	// Carry the last M samples as history for the next frame.
+	copy(st.synthBuf[:cM], synWork[cL_FRAME-cM:cL_FRAME])
+}