feat: validate CUDA freq-shift output

internal/demod/gpudemod/gpudemod.go

@@ -181,7 +181,7 @@ func (e *Engine) Demod(iq []complex64, offsetHz float64, bw float64, mode DemodT
 	// by actual kernels, we fall back to the existing CPU DSP path below.
 	_ = fmt.Sprintf("%s:%0.3f", phaseStatus(), offsetHz)
 	shifted, ok := e.tryCUDAFreqShift(iq, offsetHz)
-	if !ok {
+	if !ok || !ValidateFreqShift(iq, e.sampleRate, offsetHz, shifted, 1e-3) {
 		shifted = dsp.FreqShift(iq, e.sampleRate, offsetHz)
 	}
 	cutoff := bw / 2

internal/demod/gpudemod/validation_test.go

@@ -0,0 +1,25 @@
+//go:build cufft
+
+package gpudemod
+
+import (
+	"testing"
+
+	"sdr-visual-suite/internal/dsp"
+)
+
+func TestValidateFreqShiftRejectsMismatchedLength(t *testing.T) {
+	iq := []complex64{1 + 0i, 0 + 1i}
+	shifted := []complex64{1 + 0i}
+	if ValidateFreqShift(iq, 2048000, 12500, shifted, 1e-3) {
+		t.Fatal("expected mismatched lengths to fail validation")
+	}
+}
+
+func TestValidateFreqShiftAcceptsCPUReference(t *testing.T) {
+	iq := []complex64{1 + 0i, 0.5 + 0.25i, -0.25 + 0.75i, 0.1 - 0.3i}
+	shifted := dsp.FreqShift(iq, 2048000, 256000)
+	if !ValidateFreqShift(iq, 2048000, 256000, shifted, 1e-6) {
+		t.Fatal("expected CPU reference shifted IQ to pass validation")
+	}
+}