fix(pitch_shift): re-anchor fadePos_ on freeze so gNew is continuous at the freeze frame; guard left cast; add step-detector + loop-never-freezes regression

tests/test_pitch_shift.cpp

@@ -421,8 +421,9 @@ static void testFreezeTailContinuousTone() {
         ps.prime(src.data(), w);
         ps.setShiftRatio(2.0);
         const std::size_t preFreeze = static_cast<std::size_t>(3 * w / 4 + w / 8);
+        double lastPre = 0.0;
         for (std::size_t i = 0; i < preFreeze; ++i) {
-            (void)ps.process(src[i + static_cast<std::size_t>(w)]);
+            lastPre = static_cast<double>(ps.process(src[i + static_cast<std::size_t>(w)]));
         }
         ps.freezeTail();
         std::size_t worstGap = 0, run = 0;
@@ -450,6 +451,44 @@ static void testFreezeTailContinuousTone() {
         }
         CHECK(badZeroLat == 0);
     }
+
+    // STEP-DETECTOR: freeze-transition continuity using a ramp source where tapA and tapB
+    // read values that differ by a predictable constant (≈ A * w / N), making the crossfade
+    // gain step directly visible in the output. With a ramp, the per-frame natural change is
+    // A/(N) * ratio ≈ 0.0002 per frame; the un-fixed gain step is ~0.247 * (w/N) ≈ 0.062 —
+    // roughly 300x the natural rate. A threshold of 0.02 clearly separates fixed from unfixed.
+    //
+    // The ramp also defeats correlation-alignment (all lags score equally on a linear ramp),
+    // so the splice jump of one window guarantees tapA - tapB = −A*w/N regardless of lag.
+    {
+        PitchShifter ps;
+        ps.configure(w);
+        // Ramp from 0.0 to 1.0 over 4*w frames (same buffer size as the mid-crossfade case).
+        const std::size_t rampLen = 4 * static_cast<std::size_t>(w);
+        std::vector<AudioSample> ramp(rampLen);
+        for (std::size_t i = 0; i < rampLen; ++i) {
+            ramp[i] = static_cast<AudioSample>(static_cast<double>(i) /
+                                               static_cast<double>(rampLen - 1));
+        }
+        ps.prime(ramp.data(), w);
+        ps.setShiftRatio(2.0);
+        // Drive to the deterministic mid-fade freeze point: same preFreeze offset as above.
+        const std::size_t preFreeze = static_cast<std::size_t>(3 * w / 4 + w / 8);
+        double lastPre = 0.0;
+        for (std::size_t i = 0; i < preFreeze; ++i) {
+            lastPre = static_cast<double>(
+                ps.process(ramp[i + static_cast<std::size_t>(w)]));
+        }
+        ps.freezeTail();
+        // First frozen frame — if gNew steps at the freeze boundary the output jumps by
+        // ~deltaGain * (tapA - tapB) ≈ 0.247 * 0.25 = 0.062.
+        const double firstFrozen = static_cast<double>(ps.process(0.0f));
+        CHECK(std::isfinite(firstFrozen));
+        // Natural per-frame ramp advance at ratio 2 ≈ 2/(4*w - 1) ≈ 0.0002; the un-fixed
+        // step is ~0.062. Threshold 0.02 is 100x the natural rate but well below the step.
+        const double transitionStep = std::fabs(firstFrozen - lastPre);
+        CHECK(transitionStep < 0.02);
+    }
 }
 
 int main() {

tests/test_sampler_core.cpp

@@ -1318,6 +1318,16 @@ static void testPreserveGateStereoLoopComposes() {
     }
     CHECK(maxL > 0.05);
     CHECK(maxR > 0.02);  // R present (half amplitude), distinct from L -> stereo preserved
+    // Loop-never-freezes regression: the Preserve voice must NOT spuriously freeze when the
+    // read tap hits the sample end and the loop wraps it back. A spuriously frozen voice
+    // stops writing to the ring and the looped tail would go silent past the sample end.
+    // Check a block far past the sample end (sample = 400 frames; window = 512; well past
+    // any single-pass tail region) to catch any wrap-before-exhaustion ordering error.
+    double maxLFar = 0.0;
+    for (std::size_t i = 1800; i < 2000; ++i) {
+        if (std::fabs(left[i]) > maxLFar) maxLFar = std::fabs(left[i]);
+    }
+    CHECK(maxLFar > 0.05);  // still alive at frame 1800 (4.5× the 400-frame sample length)
 }
 
 // --- Preserve voice cap: a Preserve note-on past the cap is dropped; Varispeed unaffected. ---