Add an Oberheim-SEM morph law to the SVF filter: HP->notch->LP alongside HP->BP->LP, selected at prepare() time, free on the per-sample path

CMakeLists.txt

@@ -995,7 +995,8 @@ target_link_libraries(curve_popup PUBLIC editor_geometry)
 add_library(master_gain STATIC src/core/instrument/engine/master_gain.cpp)
 target_include_directories(master_gain PUBLIC src)
 
-# filter — the per-voice TPT/SVF with a continuous HP->BP->LP morph and an in-loop drive stage.
+# filter — the per-voice TPT/SVF with a continuous morph (HP->BP->LP or HP->notch->LP, selected
+# at prepare() time) and an in-loop drive stage.
 # The Cortex-M4 source's virtual FilterBase/Filter/Biquad hierarchy dispatched per channel per
 # sample, which the per-voice per-sample path forbids, so none of it came across. Control
 # mapping, SVF coefficients, morph weights, and the filter type each get their own file;
@@ -1108,10 +1109,29 @@ add_executable(master_gain_tests tests/test_master_gain.cpp)
 target_link_libraries(master_gain_tests PRIVATE master_gain)
 add_test(NAME master_gain_tests COMMAND master_gain_tests)
 
-# filter: the per-voice resonant filter. Pins the SVF coefficients against an independent
+# filter: four targets along the module's own seams, so each asserts one domain.
-# derivation, asserts the cutoff/Q control mappings at their anchors, holds the morph endpoints
+# filter_params_tests — the rate-free control mappings (cutoff/Q/drive) and their inverses.
-# to the analytic 2-pole targets, and measures rate/level invariance and drive stability by
+# filter_morph_tests — the pure morph-weight algebra under both morph laws; no DSP is run.
-# driving real sines. NEITHER SDK.
+# filter_state_tests — numerical stability, the denormal flush, and the state lifecycle.
+# filter_tests — the frequency response: pins the SVF coefficients against an independent
+#   derivation, holds the morph endpoints to the analytic 2-pole targets, and measures the
+#   HP-BP-LP corner flatness, the HP-notch-LP null, and rate/level invariance and drive
+#   stability by driving real sines. The seams above were chosen so this file alone owns the
+#   analytic reference and the steady-state gain measurement — a forked copy of a measurement
+#   reference is a worse defect than a long file, which is why it sits over the ~600-line bar.
+# NEITHER SDK.
+add_executable(filter_params_tests tests/test_filter_params.cpp)
+target_link_libraries(filter_params_tests PRIVATE filter)
+add_test(NAME filter_params_tests COMMAND filter_params_tests)
+
+add_executable(filter_morph_tests tests/test_filter_morph.cpp)
+target_link_libraries(filter_morph_tests PRIVATE filter)
+add_test(NAME filter_morph_tests COMMAND filter_morph_tests)
+
+add_executable(filter_state_tests tests/test_filter_state.cpp)
+target_link_libraries(filter_state_tests PRIVATE filter)
+add_test(NAME filter_state_tests COMMAND filter_state_tests)
+
 add_executable(filter_tests tests/test_filter.cpp)
 target_link_libraries(filter_tests PRIVATE filter)
 add_test(NAME filter_tests COMMAND filter_tests)

src/core/instrument/engine/filter/CLAUDE.md

@@ -3,7 +3,8 @@
 ## Scope
 
 The pure per-voice filter a sounding voice runs: a Zavalishin TPT/SVF with a continuous
-HP→BP→LP morph and a drive stage. No REAPER, no VST3, no allocation, no I/O. Everything
+morph under one of two laws — HP→BP→LP or HP→notch→LP — and a drive stage. No REAPER, no
+VST3, no allocation, no I/O. Everything
 here lives in `reasampler::instrument::engine::filter`, nested per the
 directory-mirrors-namespace convention — this keeps `FilterSettings` and friends out of
 `reasampler::instrument::engine` proper, where `zone_params.h` lives, since this module has
@@ -14,8 +15,9 @@ responsibility each:
   drive depth, plus the exact inverses for cutoff and Q.
 - `filter_coeffs` — the DSP domain: `SvfCoeffs` and the TPT coefficient solve from
   (cutoff Hz, Q, sample rate).
-- `filter_morph` — the morph domain: normalized position → per-tap weights, and the fold of
+- `filter_morph` — the morph domain: `MorphLaw`, normalized position → per-tap weights under
-  those weights into the three multipliers the kernel applies.
+  the selected law, and the fold of those weights into the three multipliers the kernel
+  applies.
 - `filter_saturate` — `softLimit`, the drive stage's shaper. Header-only inline; it sits
   inside the per-sample recursion.
 - `voice_filter` — `FilterSettings` and `VoiceFilter`, the concrete per-voice type.
@@ -63,21 +65,50 @@ now an explicit user-controlled stage instead of an emergent side effect.
 ### The morph is a blend of taps, never a coefficient switch
 
 An SVF produces high, band, and low from the same state, which is the reason this topology
-was chosen. `FilterMode` as a discrete enum is retired. HP at 0.0, BP at 0.5, LP at 1.0,
+was chosen. `FilterMode` as a discrete enum is retired. HP at 0.0, LP at 1.0, continuous
-continuous throughout, and the three endpoints are exact.
+throughout, and both endpoints are exact under either law — only the centre differs.
 
-The crossfade is **equal-power between adjacent taps**, and both halves of that are forced
+The crossfade is **equal-power** in both laws, and that is forced by the topology rather
-by the topology rather than picked by ear:
+than picked by ear. At the corner the taps are `HP = jQ`, `BP = Q`, `LP = -jQ` — adjacent
+taps in exact quadrature and HP/LP in exact antiphase, relationships the bilinear transform
+preserves exactly at the prewarped corner. A `cos`/`sin` pair therefore holds the crossfaded
+power at unity across the whole sweep; a linear crossfade of a quadrature pair would sag to
+`1/sqrt(2)` mid-leg, a 3 dB hole that reads as a defect rather than as character.
 
-- At the corner the taps are `HP = jQ`, `BP = Q`, `LP = -jQ` — adjacent taps in exact
+### The two morph laws, and why only one of them has a flat corner
-  quadrature, which the bilinear transform preserves exactly at the prewarped corner. A
+
-  `cos`/`sin` pair therefore holds the corner magnitude at exactly `Q*sqrt(cos² + sin²) = Q`
+`MorphLaw` is a two-value selector on `FilterSettings`, **defaulting to `HighBandLow`** —
-  at every morph position. A linear crossfade of a quadrature pair would sag to `Q/sqrt(2)`
+that is the reviewed-and-measured law, and it is enumerator 0 so a zero-initialized or absent
-  mid-leg — a 3 dB hole that reads as a defect, not as character.
+persisted field lands on it rather than on the SEM leg.
-- **Adjacent only.** HP and LP are exactly antiphase at the corner, so any law giving both
+
-  simultaneous weight cancels there and cuts a notch. That notch is the Oberheim SEM's
+- **`HighBandLow` (HP→BP→LP, the default).** Two equal-power legs crossfading **adjacent taps
-  centre tap. This control's centre is a band-pass, per the explicit HP/BP/LP enumeration —
+  only**, BP at the centre. Because adjacent taps are in quadrature, the corner magnitude is
-  do not "simplify" the two legs into one three-way weighting, which silently builds the SEM.
+  algebraically `Q*sqrt(cos² + sin²) = Q` at every position — measured flat to 4e-6 across 65
+  positions. **That flatness guarantee is specific to this law.** Do not weaken the assertion
+  that pins it in order to accommodate the other law.
+- **`HighNotchLow` (HP→notch→LP, the Oberheim SEM).** One equal-power crossfade weighting HP
+  and LP **together** across the whole sweep, `bp == 0` throughout. The notch is not tuned in:
+  HP and LP sit at exactly +90° and −90° at the corner, so equal weights cancel there by
+  construction. Here the corner magnitude deliberately goes to **zero** at the centre —
+  measured worst case −88 dB across every rate/cutoff/Q, typically −110 to −145 dB. The fold
+  makes that structural rather than a runtime near-miss: `m2 = lp - hp` is **exactly** `0.0f`
+  at the centre, because `cos` and `sin` of π/4 differ by about an ulp of *double*, nine
+  orders below float's spacing there, so they narrow to one float.
+
+SEM's zero is at the **notch frequency**, not a broadband level sag — off the corner the pair
+is still equal-power, so neither law's legs dip. Measuring that requires dividing by each
+tap's own analytic response first: at `Q = 0.1` a 2-pole approaches its passband so slowly
+that the pure low tap still reads 0.896 at 50 Hz, and a raw reading would report a 20% "sag"
+that is the Q, not the morph.
+
+**The toggle is free on the hot path, and must stay that way.** `morphWeights` runs at
+`prepare()` cadence; the law is consumed there and nowhere else. The kernel, `svfCoeffs`, and
+`morphMix`'s fold are identical between the laws — all a law selects is three floats the
+kernel was already multiplying by. Verified at the machine-code level, not by inspection: the
+same TU compiled `/O2` against the pre-toggle and post-toggle headers emits byte-identical
+assembly for `process()` and `processFrame()`. `VoiceFilter` gained no member and `process()`
+gained no branch. A design that puts the law selector inside the per-sample path is wrong —
+rework it rather than paying for it.
 
 ### Drive is a contraction inside the loop, which is what makes it unconditionally stable
 
@@ -183,6 +214,11 @@ topology.
 - **The morph endpoints are asserted on the folded mix, exactly.** `morphWeights` snaps the
   leg endpoints instead of trusting `cos`/`sin` to land on 0 and 1, which they miss by ~1e-17
   — enough to leave a -324 dB neighbour tap in what is specified as a pure response.
+- **A NaN morph position falls back per law, not to one shared value.** Every comparison
+  against NaN is false, so it clamps to neither endpoint: `HighBandLow` lands on pure
+  band-pass, `HighNotchLow` on pure high-pass, since it has no band tap to land on.
+- **Measuring a null needs a ring-time-adequate settle window.** At `Q = 10` the leftover
+  transient alone reads as −52 dB after 0.15 s and would be mistaken for the noise floor.
 - **No call site yet.** Wiring the filter into the voice path is a separate track; nothing
   in `sampler_core` references this module today.
 - **Decay to the denormal floor is a fixed wall-clock time, not a sample count.** A test

src/core/instrument/engine/filter/filter_morph.cpp

@@ -23,10 +23,20 @@ Pair equalPower(double t) {
 
 }  // namespace
 
-MorphWeights morphWeights(float norm) {
+MorphWeights morphWeights(float norm, MorphLaw law) {
     const double n = norm < 0.0 ? 0.0 : (norm > 1.0 ? 1.0 : static_cast<double>(norm));
 
     MorphWeights w;
+    if (law == MorphLaw::HighNotchLow) {
+        // ONE crossfade across the whole sweep rather than two legs, so HP and LP carry weight
+        // together everywhere between the endpoints and are equal at the centre.
+        const Pair p = equalPower(n);
+        w.hp = static_cast<float>(p.a);
+        w.bp = 0.0f;
+        w.lp = static_cast<float>(p.b);
+        return w;
+    }
+
     if (n <= 0.5) {
         const Pair p = equalPower(2.0 * n);  // HP -> BP
         w.hp = static_cast<float>(p.a);

src/core/instrument/engine/filter/filter_morph.h

@@ -1,34 +1,55 @@
-// filter_morph.h — the continuous HP -> BP -> LP morph: normalized position to tap weights,
+// filter_morph.h — the continuous morph: normalized position to tap weights under one of two
-// and the fold of those weights into the three multipliers the kernel actually applies. An SVF
+// laws, and the fold of those weights into the three multipliers the kernel actually applies.
-// produces all three taps from one state, so the morph is a blend, never a coefficient switch.
+// An SVF produces all three taps from one state, so the morph is a blend, never a coefficient
+// switch. Weights are computed at prepare() cadence; the law never reaches the per-sample path.
 
 #pragma once
 
 namespace reasampler::instrument::engine::filter {
 
-// Weight on each SVF tap. Exactly one of hp/lp is nonzero at a time — the morph crossfades
+// Which shape the sweep traces between its two fixed endpoints. This selects CHARACTER, not
-// between ADJACENT taps only, never HP against LP.
+// topology — same SVF, same coefficients, same kernel under either law; only the centre differs.
+//
+// HighBandLow is enumerator 0 deliberately: a zero-initialized or absent persisted field then
+// lands on the default rather than on the SEM leg.
+enum class MorphLaw {
+    // HP -> BP -> LP. Crossfades ADJACENT taps only, so the corner magnitude is flat at Q the
+    // whole way across. The default.
+    HighBandLow,
+    // HP -> notch -> LP, the Oberheim SEM. One crossfade weighting HP and LP together, bp == 0
+    // throughout; the notch falls out of the antiphase cancellation rather than being tuned in.
+    HighNotchLow,
+};
+
+// Weight on each SVF tap. Under HighBandLow exactly one of hp/lp is nonzero at a time — that law
+// crossfades adjacent taps only, never HP against LP. Under HighNotchLow bp is always zero and
+// hp/lp carry weight together, which is precisely what cuts the notch.
 struct MorphWeights {
     float hp = 0.0f;
     float bp = 0.0f;
     float lp = 1.0f;
 };
 
-// HP at 0.0, BP at 0.5, LP at 1.0. Out-of-range norm clamps to the endpoints; NaN clamps to
+// HP at 0.0, LP at 1.0 under BOTH laws; the centre is a band-pass under HighBandLow and a notch
-// neither and lands on pure band-pass instead (every comparison against it is false).
+// under HighNotchLow. Out-of-range norm clamps to the endpoints; NaN clamps to neither (every
+// comparison against it is false) and lands on the law's degenerate — pure band-pass under
+// HighBandLow, pure high-pass under HighNotchLow, which has no band tap to land on.
 //
-// Equal-power (cos/sin) rather than linear, and that choice is forced by the topology rather
+// Equal-power (cos/sin) in both laws rather than linear, and that choice is forced by the
-// than picked by ear. At the corner frequency the three taps are HP = jQ, BP = Q, LP = -jQ, so
+// topology rather than picked by ear. At the corner frequency the three taps are HP = jQ,
-// adjacent taps are in exact QUADRATURE there (and the bilinear transform preserves that exactly
+// BP = Q, LP = -jQ, so ADJACENT taps are in exact QUADRATURE there (and the bilinear transform
-// at the prewarped corner). Under a cos/sin pair the corner magnitude is therefore
+// preserves that exactly at the prewarped corner). Under HighBandLow's cos/sin pair the corner
-// Q*sqrt(cos^2 + sin^2) = Q at every morph position — algebraically flat across the whole sweep.
+// magnitude is therefore Q*sqrt(cos^2 + sin^2) = Q at every morph position — algebraically flat
-// A linear crossfade of the same quadrature pair would sag to Q/sqrt(2), a 3 dB hole mid-leg.
+// across the whole sweep. A linear crossfade of the same quadrature pair would sag to Q/sqrt(2),
+// a 3 dB hole mid-leg.
 //
-// Crossfading adjacent taps only is the other half of it: HP and LP are exactly ANTIPHASE at the
+// HP and LP are exactly ANTIPHASE at the corner (+90 and -90 degrees), so a law giving both
-// corner, so any law giving both simultaneous weight cancels there and cuts a notch. That notch
+// simultaneous weight cancels there. HighBandLow avoids that by staying adjacent; HighNotchLow
-// is the Oberheim SEM's center tap; this control's center is a band-pass, per the explicit
+// uses it — one equal-power crossfade of HP against LP over the whole sweep puts equal weights
-// HP/BP/LP enumeration.
+// at the centre and the null is exact by construction, not tuned. That is why the corner-flat-at-Q
-MorphWeights morphWeights(float norm);
+// guarantee is specific to HighBandLow: on the SEM leg the corner magnitude deliberately goes to
+// zero at the centre. Equal power still holds off the notch frequency, so neither law's legs sag.
+MorphWeights morphWeights(float norm, MorphLaw law);
 
 // The kernel applies out = m0*v0 + m1*v1 + m2*v2, where v0 is the input and v1/v2 are the SVF's
 // band and low outputs. Folding hp = v0 - k*v1 - v2 into the weights here keeps the per-sample

src/core/instrument/engine/filter/voice_filter.cpp

@@ -6,7 +6,7 @@ void VoiceFilter::prepare(const FilterSettings& settings, double sampleRate) {
     coeffs_ = svfCoeffs(filterCutoffHzFromNorm(settings.cutoffNorm),
                         filterQFromNorm(settings.resonanceNorm), sampleRate);
     if (sampleRate > 0.0) {
-        mix_ = morphMix(morphWeights(settings.morphNorm), coeffs_.k);
+        mix_ = morphMix(morphWeights(settings.morphNorm, settings.morphLaw), coeffs_.k);
     } else {
         // Bypass: a1=1, a2=a3=0 makes both state updates the exact identity, and bypassMix()
         // reads only the input, never the state -- so clearing here is audibly free (the state

src/core/instrument/engine/filter/voice_filter.h

@@ -15,11 +15,14 @@
 namespace reasampler::instrument::engine::filter {
 
 // Normalized control positions, as the editor moves them and the persisted state carries them.
+// morphLaw is the one discrete control here — a two-value selector, not a normalized position —
+// because its two values are characters to choose between, not points on a continuum.
 struct FilterSettings {
     float cutoffNorm = 1.0f;
     float resonanceNorm = 0.0f;
-    float morphNorm = 1.0f;  // 0 = high-pass, 0.5 = band-pass, 1 = low-pass
+    float morphNorm = 1.0f;  // 0 = high-pass, 1 = low-pass; the centre is set by morphLaw
     float driveNorm = 0.0f;
+    MorphLaw morphLaw = MorphLaw::HighBandLow;
 };
 
 // Below this the recursion has decayed past -600 dB. Flushing keeps the state out of the

tests/test_filter.cpp

@@ -1,8 +1,11 @@
-// Standalone tests for the per-voice TPT/SVF filter — no VST3, no REAPER, no framework. Same
+// Standalone tests for the RUNNING per-voice TPT/SVF filter — no VST3, no REAPER, no framework.
-// fast assert loop as the sibling pure tests. The coefficient pins are literals so a refactor
+// Same fast assert loop as the sibling pure tests. The coefficient pins are literals so a
-// that changes the DSP fails loudly; they are cross-checked in-test against a derivation that
+// refactor that changes the DSP fails loudly; they are cross-checked in-test against a derivation
-// shares no code with the implementation, and the responses against the analog 2-pole prototype
+// that shares no code with the implementation, and the responses against the analog 2-pole
-// evaluated at the bilinear-warped frequency.
+// prototype evaluated at the bilinear-warped frequency. Sibling targets own the neighbouring
+// domains: test_filter_params.cpp the control mappings, test_filter_morph.cpp the pure morph-weight
+// algebra, test_filter_state.cpp the numerical/state behaviour. This file owns the analytic
+// reference and the steady-state gain measurement, and everything here uses them.
 
 #include "../src/core/instrument/engine/filter/filter_coeffs.h"
 #include "../src/core/instrument/engine/filter/filter_morph.h"
@@ -10,11 +13,9 @@
 #include "../src/core/instrument/engine/filter/filter_saturate.h"
 #include "../src/core/instrument/engine/filter/voice_filter.h"
 
-#include <cfloat>
 #include <cmath>
 #include <cstdio>
 #include <initializer_list>
-#include <limits>
 
 using namespace reasampler::instrument::engine::filter;
 
@@ -28,11 +29,18 @@ static int g_fail = 0;
 
 static constexpr double kPi = 3.14159265358979323846;
 
-// Morph positions of the three pure taps.
+// Morph positions. The endpoints are the same pure taps under both laws; only the centre differs
+// — a band-pass under HighBandLow, a notch under HighNotchLow.
 static constexpr float kHighPass = 0.0f;
 static constexpr float kBandPass = 0.5f;
+static constexpr float kCentre = 0.5f;
 static constexpr float kLowPass = 1.0f;
 
+static const MorphLaw kBothLaws[] = {MorphLaw::HighBandLow, MorphLaw::HighNotchLow};
+static const char* lawName(MorphLaw law) {
+    return law == MorphLaw::HighBandLow ? "HP-BP-LP" : "HP-notch-LP";
+}
+
 // The rates the invariance claims are made over.
 static const double kRates[] = {44100.0, 48000.0, 88200.0, 96000.0, 192000.0};
 static constexpr int kRateCount = 5;
@@ -75,101 +83,9 @@ static double measuredGain(const FilterSettings& fs, double sr, double freq, dou
     return std::sqrt(sumSq / measure) / (amp / std::sqrt(2.0));
 }
 
-static FilterSettings at(double fcHz, float res, float morph, float drive = 0.0f) {
+static FilterSettings at(double fcHz, float res, float morph, float drive = 0.0f,
-    return {filterNormFromCutoffHz(static_cast<float>(fcHz)), res, morph, drive};
+                         MorphLaw law = MorphLaw::HighBandLow) {
-}
+    return {filterNormFromCutoffHz(static_cast<float>(fcHz)), res, morph, drive, law};
-
-// ---------------------------------------------------------------------------
-// Control mappings (carried over — the cutoff and Q laws are unchanged)
-// ---------------------------------------------------------------------------
-
-static void testCutoffMapsThreeDecadesLogarithmically() {
-    CHECK_NEAR(filterCutoffHzFromNorm(0.0f), 20.0, 1e-3);
-    CHECK_NEAR(filterCutoffHzFromNorm(1.0f), 20000.0, 1e-2);
-
-    // Exactly three decades, so the decade midpoints land on round numbers.
-    CHECK_NEAR(filterCutoffHzFromNorm(1.0f / 3.0f), 200.0, 1e-3);
-    CHECK_NEAR(filterCutoffHzFromNorm(2.0f / 3.0f), 2000.0, 1e-2);
-
-    // Half-decade steps confirm the sweep is log, not linear.
-    CHECK_NEAR(filterCutoffHzFromNorm(1.0f / 6.0f), 20.0 * std::sqrt(10.0), 1e-3);
-    CHECK_NEAR(filterCutoffHzFromNorm(0.5f), 20.0 * std::sqrt(1000.0), 1e-2);
-
-    CHECK_NEAR(filterCutoffHzFromNorm(-1.0f), 20.0, 1e-3);
-    CHECK_NEAR(filterCutoffHzFromNorm(2.0f), 20000.0, 1e-2);
-}
-
-static void testCutoffNormRoundTrips() {
-    for (int i = 0; i <= 20; ++i) {
-        const float n = static_cast<float>(i) / 20.0f;
-        CHECK_NEAR(filterNormFromCutoffHz(filterCutoffHzFromNorm(n)), n, 1e-6);
-    }
-    CHECK_NEAR(filterNormFromCutoffHz(200.0f), 1.0 / 3.0, 1e-6);
-    CHECK(filterNormFromCutoffHz(1.0f) == 0.0f);
-    CHECK(filterNormFromCutoffHz(48000.0f) == 1.0f);
-}
-
-static void testQSpansPointOneToTenWithRootTwoAtCenter() {
-    CHECK_NEAR(filterQFromNorm(0.0f), 0.1, 1e-6);
-    CHECK_NEAR(filterQFromNorm(0.5f), std::sqrt(2.0), 1e-5);
-    CHECK_NEAR(filterQFromNorm(1.0f), 10.0, 1e-4);
-
-    CHECK_NEAR(filterQFromNorm(-1.0f), 0.1, 1e-6);
-    CHECK_NEAR(filterQFromNorm(2.0f), 10.0, 1e-4);
-
-    // Pins the single quadratic-in-log-Q curve at two interior points, derived independently by
-    // solving log Q = a + b*n + c*n^2 through the three anchors above rather than read out of
-    // the implementation. A two-spliced-log-segments curve (log-linear on each half, the design
-    // this module doc explicitly rejects for its center-detent slope kink) would give 0.376 and
-    // 3.761 here instead — both comfortably outside this tolerance.
-    {
-        const double lo = std::log(static_cast<double>(kFilterQMin));
-        const double mid = std::log(static_cast<double>(kFilterQCenter));
-        const double hi = std::log(static_cast<double>(kFilterQMax));
-        const double c = 2.0 * lo + 2.0 * hi - 4.0 * mid;
-        const double b = hi - lo - c;
-        const double a = lo;
-        auto qLaw = [&](double n) { return std::exp(a + b * n + c * n * n); };
-        CHECK_NEAR(filterQFromNorm(0.25f), qLaw(0.25), 1e-5);
-        CHECK_NEAR(filterQFromNorm(0.75f), qLaw(0.75), 1e-5);
-    }
-
-    // Strictly monotonic across the whole travel — no fold-back from the quadratic term.
-    float prev = -1.0f;
-    for (int i = 0; i <= 1000; ++i) {
-        const float q = filterQFromNorm(static_cast<float>(i) / 1000.0f);
-        CHECK(q > prev);
-        prev = q;
-    }
-}
-
-static void testQNormRoundTrips() {
-    for (int i = 0; i <= 20; ++i) {
-        const float n = static_cast<float>(i) / 20.0f;
-        CHECK_NEAR(filterNormFromQ(filterQFromNorm(n)), n, 1e-5);
-    }
-    CHECK_NEAR(filterNormFromQ(static_cast<float>(std::sqrt(2.0))), 0.5, 1e-5);
-    CHECK(filterNormFromQ(0.0f) == 0.0f);
-    CHECK(filterNormFromQ(1000.0f) == 1.0f);
-}
-
-static void testDriveDepthIsZeroAtRestAndRisesMonotonically() {
-    // Exactly zero, not nearly: the limiter is the identity only at depth 0.
-    CHECK(filterDriveDepthFromNorm(0.0f) == 0.0f);
-    CHECK(filterDriveDepthFromNorm(-1.0f) == 0.0f);
-    CHECK_NEAR(filterDriveDepthFromNorm(1.0f), kFilterDriveDepthMax, 1e-6);
-    CHECK_NEAR(filterDriveDepthFromNorm(2.0f), kFilterDriveDepthMax, 1e-6);
-
-    // Pins the SQUARE law at an interior point, not just the anchors: a linear law would give
-    // kFilterDriveDepthMax/2 (2.0) here, not kFilterDriveDepthMax/4 (1.0).
-    CHECK_NEAR(filterDriveDepthFromNorm(0.5f), kFilterDriveDepthMax * 0.25, 1e-6);
-
-    float prev = -1.0f;
-    for (int i = 0; i <= 100; ++i) {
-        const float d = filterDriveDepthFromNorm(static_cast<float>(i) / 100.0f);
-        CHECK(d > prev);
-        prev = d;
-    }
 }
 
 // ---------------------------------------------------------------------------
@@ -253,69 +169,25 @@ static void testNonPositiveRatePassesSignalThroughAtEveryMorph() {
 }
 
 // ---------------------------------------------------------------------------
-// Morph
+// Morph — measured, under both laws
 // ---------------------------------------------------------------------------
 
-// The endpoints are pure taps EXACTLY, not to within a rounding of cos/sin. Asserted on the
+// The endpoints are exact 2-pole HP and LP under BOTH laws; only the centre is law-specific, so
-// folded mix, where "pure" is an exact statement about three floats.
+// the centre is asserted here only for the law that has a pure tap there.
-static void testMorphEndpointMixesAreExactlyPureTaps() {
-    const float k = 1.0f / filterQFromNorm(0.5f);
-
-    const MorphMix hp = morphMix(morphWeights(kHighPass), k);
-    CHECK(hp.m0 == 1.0f && hp.m1 == -k && hp.m2 == -1.0f);  // v0 - k*v1 - v2
-
-    const MorphMix bp = morphMix(morphWeights(kBandPass), k);
-    CHECK(bp.m0 == 0.0f && bp.m1 == 1.0f && bp.m2 == 0.0f);  // v1
-
-    const MorphMix lp = morphMix(morphWeights(kLowPass), k);
-    CHECK(lp.m0 == 0.0f && lp.m1 == 0.0f && lp.m2 == 1.0f);  // v2
-
-    // Out-of-range clamps to the endpoints rather than extrapolating.
-    CHECK(morphWeights(-1.0f).hp == 1.0f);
-    CHECK(morphWeights(2.0f).lp == 1.0f);
-
-    // NaN clamps to neither endpoint (every comparison against it is false) and lands on pure
-    // band-pass instead -- no crash, a sane fallback rather than an extrapolation.
-    const MorphWeights nanW = morphWeights(std::numeric_limits<float>::quiet_NaN());
-    CHECK(nanW.hp == 0.0f && nanW.bp == 1.0f && nanW.lp == 0.0f);
-}
-
-// Pins the cos/sin curve at an interior point, not just the endpoints and the quadrature
-// identity (hp^2+bp^2+lp^2=1, which any equal-power reparameterization would also satisfy).
-// theta=0.5*pi*t^2 (quadratic in the leg fraction, still equal-power, still exact at both
-// ends) would give hp=0.9239/bp=0.3827 here instead of the cos/sin pair's 0.7071/0.7071.
-static void testMorphInteriorPointMatchesCosSinNotAnAlternateEqualPowerCurve() {
-    const MorphWeights w = morphWeights(0.25f);  // HP->BP leg, t = 2*0.25 = 0.5
-    const double theta = 0.5 * kPi * 0.5;
-    CHECK_NEAR(w.hp, std::cos(theta), 1e-6);
-    CHECK_NEAR(w.bp, std::sin(theta), 1e-6);
-    CHECK(w.lp == 0.0f);
-}
-
-// HP and LP never carry weight at the same time. That is what keeps the centre a band-pass
-// instead of the Oberheim SEM's notch: the two are antiphase at the corner and would cancel.
-static void testMorphNeverBlendsHighAgainstLowPass() {
-    for (int i = 0; i <= 200; ++i) {
-        const MorphWeights w = morphWeights(static_cast<float>(i) / 200.0f);
-        CHECK(w.hp == 0.0f || w.lp == 0.0f);
-        CHECK(w.hp >= 0.0f && w.bp >= 0.0f && w.lp >= 0.0f);
-        // Equal power: the active pair sums in quadrature to unity.
-        CHECK_NEAR(w.hp * w.hp + w.bp * w.bp + w.lp * w.lp, 1.0, 1e-6);
-    }
-}
-
 static void testMorphEndpointsMatchTheAnalyticTwoPoleTargets() {
     const double sr = 48000.0, fc = 1000.0;
+    for (MorphLaw law : kBothLaws) {
         for (float res : {0.0f, 0.5f, 1.0f}) {
             const double q = filterQFromNorm(res);
             for (float morph : {kHighPass, kBandPass, kLowPass}) {
+                if (morph == kBandPass && law != MorphLaw::HighBandLow) continue;
                 for (double f : {125.0, 500.0, 1000.0, 2000.0, 8000.0}) {
-                const double got = measuredGain(at(fc, res, morph), sr, f);
+                    const double got = measuredGain(at(fc, res, morph, 0.0f, law), sr, f);
                     const double want = analyticMag(morph, f, fc, q, sr);
                     if (!(std::fabs(got / want - 1.0) <= kAgreement)) {
-                    std::printf("FAIL line %d: morph %.1f res %.1f at %.0f Hz: %.6f vs analytic "
+                        std::printf("FAIL line %d: %s morph %.1f res %.1f at %.0f Hz: %.6f vs "
-                                "%.6f (%.3f%%)\n",
+                                    "analytic %.6f (%.3f%%)\n",
-                                __LINE__, morph, res, f, got, want,
+                                    __LINE__, lawName(law), morph, res, f, got, want,
                                     (got / want - 1.0) * 100.0);
                         ++g_fail;
                     }
@@ -323,18 +195,21 @@ static void testMorphEndpointsMatchTheAnalyticTwoPoleTargets() {
             }
         }
     }
+}
 
-// The reason the blend is equal-power rather than linear. At the corner the three taps are
+// LAW-SPECIFIC, and deliberately not generalized: this guarantee belongs to HighBandLow alone.
-// HP = jQ, BP = Q, LP = -jQ — adjacent taps in exact quadrature — so a cos/sin pair holds the
+// At the corner the three taps are HP = jQ, BP = Q, LP = -jQ — ADJACENT taps in exact quadrature
-// corner magnitude at exactly Q the whole way across. A linear crossfade would sag to Q/sqrt(2)
+// — so a cos/sin pair holds the corner magnitude at exactly Q the whole way across. A linear
-// mid-leg, a 3 dB hole that would read as a defect rather than as character.
+// crossfade would sag to Q/sqrt(2) mid-leg, a 3 dB hole that would read as a defect rather than
-static void testCornerMagnitudeIsFlatAcrossTheWholeMorphSweep() {
+// as character. HighNotchLow deliberately violates this (its corner magnitude goes to zero at the
+// centre); weakening this assertion to accommodate that law would throw the guarantee away.
+static void testCornerMagnitudeIsFlatAtQAcrossTheHighBandLowSweep() {
     const double sr = 48000.0, fc = 1000.0;
     for (float res : {0.0f, 0.5f, 1.0f}) {
         const double q = filterQFromNorm(res);
         for (int i = 0; i <= 16; ++i) {
             const float m = static_cast<float>(i) / 16.0f;
-            const double got = measuredGain(at(fc, res, m), sr, fc);
+            const double got = measuredGain(at(fc, res, m, 0.0f, MorphLaw::HighBandLow), sr, fc);
             if (!(std::fabs(got / q - 1.0) <= kAgreement)) {
                 std::printf("FAIL line %d: morph %.4f res %.1f corner gain %.6f, expected Q "
                             "%.6f (%.3f%%)\n",
@@ -345,29 +220,116 @@ static void testCornerMagnitudeIsFlatAcrossTheWholeMorphSweep() {
     }
 }
 
+// The SEM's centre is a genuine null, not merely a dip: the corner magnitude falls to the float
+// noise floor because HP and LP sit at exactly +90 and -90 degrees there, so equal weights cancel
+// by construction. Measured worst case across this whole grid is 3.8e-05 (-88 dB); the typical
+// figure is -110 to -145 dB. The settle window has to clear the resonator's ring-down before the
+// residual means anything — at 0.15 s and Q=10 the leftover transient alone reads as -52 dB and
+// would be mistaken for the floor.
+static void testHighNotchLowCentreIsATrueNullAtTheCorner() {
+    for (int r = 0; r < kRateCount; ++r) {
+        for (double fc : {250.0, 1000.0, 4000.0}) {
+            for (float res : {0.0f, 0.5f, 1.0f}) {
+                const double got = measuredGain(at(fc, res, kCentre, 0.0f, MorphLaw::HighNotchLow),
+                                                kRates[r], fc, 0.25, 2.0, 0.5);
+                if (!(got < 2e-4)) {
+                    std::printf("FAIL line %d: SEM notch at sr %.0f fc %.0f res %.1f is %.3e "
+                                "(%.1f dB) — not a null\n",
+                                __LINE__, kRates[r], fc, res, got,
+                                20.0 * std::log10(got + 1e-300));
+                    ++g_fail;
+                }
+            }
+        }
+    }
+}
+
+// The null sits AT the cutoff, not merely somewhere nearby: the response falls monotonically into
+// fc from both sides and is orders of magnitude below its own immediate neighbours. At fc=1 kHz,
+// +/-5% off the notch already reads -20 dB while the notch itself reads -127 dB.
+static void testHighNotchLowNullIsLocatedAtTheCutoff() {
+    const double sr = 48000.0, fc = 1000.0;
+    for (float res : {0.0f, 0.5f, 1.0f}) {
+        const FilterSettings fs = at(fc, res, kCentre, 0.0f, MorphLaw::HighNotchLow);
+        const double below[] = {0.5, 0.8, 0.95};
+        double prev = 1e30;
+        for (double ratio : below) {
+            const double got = measuredGain(fs, sr, fc * ratio);
+            CHECK(got < prev);
+            prev = got;
+        }
+        const double atCorner = measuredGain(fs, sr, fc, 0.25, 2.0, 0.5);
+        CHECK(atCorner < prev);
+
+        prev = atCorner;
+        for (double ratio : {1.05, 1.25, 2.0}) {
+            const double got = measuredGain(fs, sr, fc * ratio);
+            CHECK(got > prev);
+            prev = got;
+        }
+        // Against its own immediate neighbours, so this is a null rather than a broad scoop.
+        CHECK(atCorner < 1e-3 * measuredGain(fs, sr, fc * 0.95));
+    }
+}
+
+// The SEM's zero is AT the notch frequency, not a broadband level sag: away from the corner the
+// two taps are still an equal-power pair, so the sweep holds constant power on its legs. Measured
+// deep in each tap's own passband — 50 Hz for the low tap, 20 kHz for the high tap, both far from
+// a 1 kHz corner — and divided by that tap's OWN analytic response there, so what is left is the
+// weight the law applied. That normalization is load-bearing, not cosmetic: at Q = 0.1 a 2-pole
+// approaches its passband so slowly that the pure low tap still reads 0.896 at 50 Hz, and a raw
+// reading would report a 20% "sag" that is the Q, not the morph. A LINEAR crossfade would give
+// 0.5 at the centre instead of 1.0, so this tolerance discriminates equal-power from linear
+// decisively rather than merely confirming a plausible shape.
+static void testHighNotchLowLegsHoldConstantPowerAwayFromTheNotch() {
+    const double sr = 48000.0, fc = 1000.0;
+    for (float res : {0.0f, 0.5f, 1.0f}) {
+        const double q = filterQFromNorm(res);
+        const double lowRef = analyticMag(kLowPass, 50.0, fc, q, sr);
+        const double highRef = analyticMag(kHighPass, 20000.0, fc, q, sr);
+        for (int i = 0; i <= 8; ++i) {
+            const float m = static_cast<float>(i) / 8.0f;
+            const FilterSettings fs = at(fc, res, m, 0.0f, MorphLaw::HighNotchLow);
+            const double low = measuredGain(fs, sr, 50.0) / lowRef;
+            const double high = measuredGain(fs, sr, 20000.0) / highRef;
+            const double power = low * low + high * high;
+            if (!(std::fabs(power - 1.0) <= 0.02)) {
+                std::printf("FAIL line %d: SEM morph %.3f res %.1f leg power %.6f (low %.6f, "
+                            "high %.6f) — expected 1.0\n",
+                            __LINE__, m, res, power, low, high);
+                ++g_fail;
+            }
+        }
+    }
+}
+
 // Continuity as a control, not just at the corner: no step between adjacent morph positions at
-// any fixed frequency. A coefficient switch at the centre — the thing an enum would have forced —
+// any fixed frequency, under either law. A coefficient switch at the centre — the thing an enum
-// shows up here as a jump.
+// over TOPOLOGIES would have forced — shows up here as a jump. Measured off the SEM's notch
+// frequency, since the null itself is a legitimate near-step in the response.
 static void testMorphSweepHasNoDiscontinuity() {
     const double sr = 48000.0, fc = 1000.0;
     constexpr int kSteps = 40;
+    for (MorphLaw law : kBothLaws) {
         for (float res : {0.0f, 0.5f, 1.0f}) {
             for (double f : {250.0, 1000.0, 4000.0}) {
+                if (f == fc && law == MorphLaw::HighNotchLow) continue;
                 double prev = -1.0;
                 for (int i = 0; i <= kSteps; ++i) {
                     const float m = static_cast<float>(i) / kSteps;
-                const double got = measuredGain(at(fc, res, m), sr, f);
+                    const double got = measuredGain(at(fc, res, m, 0.0f, law), sr, f);
                     if (prev >= 0.0) {
                         // Scaled by the response's own magnitude at this setting — the passband is
-                    // unity and the corner is Q, so below Q=1 the passband is what a step has to
+                        // unity and the corner is Q, so below Q=1 the passband is what a step has
-                    // be small against, not Q.
+                        // to be small against, not Q.
                         const double scale = std::fmax(1.0, filterQFromNorm(res));
-                    // One step is 1/40 of the travel; the steepest leg moves well under a tenth
+                        // One step is 1/40 of the travel; the steepest leg moves well under a
-                    // of that scale over one step (measured worst case is 0.03).
+                        // tenth of that scale over one step (measured worst case is 0.03).
                         const double jump = std::fabs(got - prev) / scale;
                         if (!(jump < 0.1)) {
-                        std::printf("FAIL line %d: morph %.4f res %.1f at %.0f Hz jumps %.4f\n",
+                            std::printf("FAIL line %d: %s morph %.4f res %.1f at %.0f Hz jumps "
-                                    __LINE__, m, res, f, jump);
+                                        "%.4f\n",
+                                        __LINE__, lawName(law), m, res, f, jump);
                             ++g_fail;
                         }
                     }
@@ -376,6 +338,44 @@ static void testMorphSweepHasNoDiscontinuity() {
             }
         }
     }
+}
+
+// The law selects a MIX, computed once per prepare(); it must not reach the coefficient solve at
+// all. Asserted bit-exactly rather than by tolerance — the cutoff, the damping term, and the
+// zero-delay-loop solution are the same floats under either law, so no cutoff/Q/rate behaviour
+// can differ between them by construction.
+static void testMorphLawDoesNotDisturbTheCoefficients() {
+    for (int r = 0; r < kRateCount; ++r) {
+        for (int ci = 0; ci <= 8; ++ci) {
+            for (float res : {0.0f, 0.5f, 1.0f}) {
+                for (int mi = 0; mi <= 4; ++mi) {
+                    VoiceFilter band, sem;
+                    const float m = mi / 4.0f;
+                    band.prepare({ci / 8.0f, res, m, 0.5f, MorphLaw::HighBandLow}, kRates[r]);
+                    sem.prepare({ci / 8.0f, res, m, 0.5f, MorphLaw::HighNotchLow}, kRates[r]);
+                    const SvfCoeffs& a = band.coeffs();
+                    const SvfCoeffs& b = sem.coeffs();
+                    CHECK(a.g == b.g && a.k == b.k);
+                    CHECK(a.a1 == b.a1 && a.a2 == b.a2 && a.a3 == b.a3);
+                }
+            }
+        }
+    }
+}
+
+// The default is the reviewed-and-measured law, not the SEM leg. The editor and any persisted-
+// state codec read this default, so a preset saved before the selector existed must still sound
+// exactly as it did — asserted on the folded mix, which is the only thing the kernel sees.
+static void testFilterSettingsDefaultsToTheHighBandLowLaw() {
+    CHECK(FilterSettings{}.morphLaw == MorphLaw::HighBandLow);
+
+    VoiceFilter defaulted, explicitLaw;
+    defaulted.prepare({0.5f, 0.5f, kCentre, 0.0f}, 48000.0);
+    explicitLaw.prepare({0.5f, 0.5f, kCentre, 0.0f, MorphLaw::HighBandLow}, 48000.0);
+    CHECK(defaulted.mix().m0 == explicitLaw.mix().m0);
+    CHECK(defaulted.mix().m1 == explicitLaw.mix().m1);
+    CHECK(defaulted.mix().m2 == explicitLaw.mix().m2);
+}
 
 // ---------------------------------------------------------------------------
 // Drive
@@ -411,10 +411,9 @@ static float nextNoise(unsigned& rng) {
            static_cast<float>(1 << 22);
 }
 
-static void testDriveZeroIsBitIdenticalToTheLinearKernel() {
+static void checkDriveZeroBitIdentity(float morph, MorphLaw law) {
-    for (float morph : {kHighPass, kBandPass, kLowPass}) {
     VoiceFilter f;
-        f.prepare(at(1000.0, 1.0f, morph, 0.0f), 48000.0);
+    f.prepare(at(1000.0, 1.0f, morph, 0.0f, law), 48000.0);
     f.reset();
     LinearKernelRef ref0{f.coeffs(), f.mix()};
     LinearKernelRef ref1{f.coeffs(), f.mix()};
@@ -443,6 +442,11 @@ static void testDriveZeroIsBitIdenticalToTheLinearKernel() {
         CHECK(frame[1] == want1);
     }
 }
+
+static void testDriveZeroIsBitIdenticalToTheLinearKernel() {
+    for (MorphLaw law : kBothLaws) {
+        for (float morph : {kHighPass, kBandPass, kLowPass}) checkDriveZeroBitIdentity(morph, law);
+    }
 }
 
 // The complaint the rewrite answers: resonance must not track how hard the sample hits the
@@ -478,18 +482,19 @@ static void testFullDriveStaysBoundedAtEveryCutoffResonanceAndRate() {
 
     for (int r = 0; r < kRateCount; ++r) {
         const double sr = kRates[r];
+        for (MorphLaw law : kBothLaws) {
             for (int ci = 0; ci <= 8; ++ci) {
                 for (int mi = 0; mi <= 4; ++mi) {
                     for (float res : {0.0f, 0.5f, 1.0f}) {
                         VoiceFilter f;
-                    f.prepare({ci / 8.0f, res, mi / 4.0f, 1.0f}, sr);
+                        f.prepare({ci / 8.0f, res, mi / 4.0f, 1.0f, law}, sr);
                         f.reset();
                         for (int i = 0; i < 4000; ++i) {
                             const float y = f.process(0, noise());
                             if (!std::isfinite(y) || std::fabs(y) > 8.0f) {
-                            std::printf("FAIL line %d: sr=%.0f cutoff=%.2f morph=%.2f res=%.1f "
+                                std::printf("FAIL line %d: %s sr=%.0f cutoff=%.2f morph=%.2f "
-                                        "full drive produced %g\n",
+                                            "res=%.1f full drive produced %g\n",
-                                        __LINE__, sr, ci / 8.0, mi / 4.0, res, y);
+                                            __LINE__, lawName(law), sr, ci / 8.0, mi / 4.0, res, y);
                                 ++g_fail;
                                 return;
                             }
@@ -499,15 +504,17 @@ static void testFullDriveStaysBoundedAtEveryCutoffResonanceAndRate() {
             }
         }
     }
+}
 
 // Full drive at full resonance with no input must still go quiet. A nonlinearity in the loop is
 // exactly where a self-oscillator would hide, and softLimit's sub-unit slope is what forbids it.
 static void testFullDriveDoesNotSelfOscillate() {
     for (int r = 0; r < kRateCount; ++r) {
         const double sr = kRates[r];
+        for (MorphLaw law : kBothLaws) {
             for (float morph : {kHighPass, kBandPass, kLowPass}) {
                 VoiceFilter f;
-            f.prepare(at(1000.0, 1.0f, morph, 1.0f), sr);
+                f.prepare(at(1000.0, 1.0f, morph, 1.0f, law), sr);
                 f.reset();
                 const int excite = static_cast<int>(sr * 0.01);
                 for (int i = 0; i < excite; ++i) {
@@ -518,6 +525,7 @@ static void testFullDriveDoesNotSelfOscillate() {
             }
         }
     }
+}
 
 // Drive has to actually do something at the top of its travel, and do it monotonically — the
 // brief's "extreme, not politely warm". Measured at the corner, where the resonance state is
@@ -621,198 +629,20 @@ static void testLowCutoffHighRateCornerHoldsTheAnalyticPeak() {
     }
 }
 
-// ---------------------------------------------------------------------------
-// Stability, denormals, and state
-// ---------------------------------------------------------------------------
-
-static void testFullRangeCutoffSweepAtAudioRateStaysBounded() {
-    unsigned rng = 0x13579bdfu;
-    auto noise = [&rng]() {
-        rng = rng * 1664525u + 1013904223u;
-        return static_cast<float>(static_cast<int>(rng >> 9) - (1 << 22)) /
-               static_cast<float>(1 << 22);
-    };
-
-    for (int r = 0; r < kRateCount; ++r) {
-        const double sr = kRates[r];
-        for (float morph : {kHighPass, kBandPass, kLowPass}) {
-            for (float res : {0.0f, 1.0f}) {
-                for (float drive : {0.0f, 1.0f}) {
-                    VoiceFilter f;
-                    f.reset();
-                    // A fixed WALL-CLOCK sweep: the same cutoff travel per second at every rate,
-                    // so the per-sample coefficient step gets no gentler as the rate rises.
-                    const int n = static_cast<int>(sr * 0.25);
-                    for (int i = 0; i < n; ++i) {
-                        const float t = static_cast<float>(i) / static_cast<float>(n - 1);
-                        f.prepare({t, res, morph, drive}, sr);
-                        const float y = f.process(0, noise());
-                        CHECK(std::isfinite(y));
-                        CHECK(std::fabs(y) < 100.0f);
-                        if (!std::isfinite(y)) return;  // stop before the log floods
-                    }
-                }
-            }
-        }
-    }
-}
-
-// The flush tests the ENVELOPE — both integrators — not one sample. ic1 and ic2 are in
-// quadrature, so a resonator swings each through zero twice a cycle; flushing on a single one
-// injects a step in phase with the resonance, which the resonance amplifies, and the filter
-// limit-cycles at the floor forever instead of going quiet. Re-verified for TPT rather than
-// assumed to carry over from the retired Direct Form I state.
-static void testStateFlushesToZeroWithoutStallingInDenormals() {
-    for (int r = 0; r < kRateCount; ++r) {
-        const double sr = kRates[r];
-        // The decay to the floor is a fixed WALL-CLOCK time, so the budget scales with the rate.
-        const int budget = static_cast<int>(sr * 0.5);
-        for (float morph : {kHighPass, kBandPass, kLowPass}) {
-            for (float drive : {0.0f, 1.0f}) {
-                VoiceFilter f;
-                f.prepare(at(1000.0, 1.0f, morph, drive), sr);
-                f.reset();
-
-                // Excite, then hard-cut to silence the way a released voice does.
-                const int excite = static_cast<int>(sr * 0.01);
-                for (int i = 0; i < excite; ++i) {
-                    f.process(0, 0.5f * static_cast<float>(std::sin(2.0 * kPi * 1000.0 * i / sr)));
-                }
-
-                int subnormalSamples = 0, silentAt = -1;
-                for (int i = 0; i < budget; ++i) {
-                    f.process(0, 0.0f);
-                    const VoiceFilter::State& s = f.state(0);
-                    if ((s.ic1 != 0.0f && std::fabs(s.ic1) < FLT_MIN) ||
-                        (s.ic2 != 0.0f && std::fabs(s.ic2) < FLT_MIN)) {
-                        ++subnormalSamples;
-                    }
-                    if (silentAt < 0 && f.isSilent()) silentAt = i;
-                }
-                // Without the flush the state grinds down through the subnormal range for
-                // thousands of samples; a stray sample or two at a zero crossing is not a stall.
-                CHECK(subnormalSamples <= 2);
-                CHECK(silentAt >= 0);
-                CHECK(silentAt < budget);
-                // And it stays silent — a flush that perturbs the loop would re-excite it.
-                for (int i = 0; i < 1000; ++i) CHECK(f.process(0, 0.0f) == 0.0f);
-                CHECK(f.isSilent());
-            }
-        }
-    }
-}
-
-// A high-pass under sustained DC must settle to zero and STAY there. Sampling only the final
-// value is not enough: a resonator swings through zero twice a cycle, so a single late sample
-// can land near zero while the envelope still rings well above it. This regressed a click train
-// on the retired topology, where flushing the FIR history discarded the pinned DC and the next
-// sample recomputed a full-amplitude step. TPT has no FIR history to discard, so the hazard is
-// structural rather than a tuning — but the assertion is cheap and pins the outcome.
-static void testHighPassSustainedDCDoesNotReRing() {
-    for (int r = 0; r < kRateCount; ++r) {
-        const double sr = kRates[r];
-        for (float drive : {0.0f, 1.0f}) {
-            VoiceFilter f;
-            f.prepare(at(1000.0, 1.0f, kHighPass, drive), sr);
-            f.reset();
-            const int settle = static_cast<int>(sr * 0.05);
-            float worstAfterSettle = 0.0f;
-            for (int i = 0; i < static_cast<int>(sr * 0.5); ++i) {
-                const float y = f.process(0, 1.0f);
-                if (i >= settle) worstAfterSettle = std::fmax(worstAfterSettle, std::fabs(y));
-            }
-            CHECK(worstAfterSettle < 1e-3f);
-        }
-    }
-}
-
-static void testImpulseResponseMatchesTheKernel() {
-    VoiceFilter f;
-    f.prepare(at(1000.0, 0.5f, kLowPass), 48000.0);
-    f.reset();
-    const SvfCoeffs c = f.coeffs();
-    // From a cleared state the first sample reduces to the coefficients alone: v1 == a2, v2 == a3.
-    CHECK_NEAR(f.process(0, 1.0f), c.a3, 1e-7);
-
-    VoiceFilter bp;
-    bp.prepare(at(1000.0, 0.5f, kBandPass), 48000.0);
-    bp.reset();
-    CHECK_NEAR(bp.process(0, 1.0f), c.a2, 1e-7);
-
-    VoiceFilter hp;
-    hp.prepare(at(1000.0, 0.5f, kHighPass), 48000.0);
-    hp.reset();
-    CHECK_NEAR(hp.process(0, 1.0f), 1.0 - c.k * c.a2 - c.a3, 1e-7);
-}
-
-static void testLowpassStepSettlesToUnityAndHighpassRejectsDC() {
-    const double sr = 48000.0;
-    VoiceFilter f;
-    f.prepare(at(1000.0, 0.0f, kLowPass), sr);
-    f.reset();
-    float y = 0.0f;
-    for (int i = 0; i < 48000; ++i) y = f.process(0, 1.0f);
-    CHECK_NEAR(y, 1.0, 1e-3);  // DC passes a lowpass at unity
-
-    VoiceFilter hp;
-    hp.prepare(at(1000.0, 0.0f, kHighPass), sr);
-    hp.reset();
-    float worstAfterSettle = 0.0f;
-    for (int i = 0; i < 48000; ++i) {
-        y = hp.process(0, 1.0f);
-        if (i >= 200) worstAfterSettle = std::fmax(worstAfterSettle, std::fabs(y));
-    }
-    CHECK(worstAfterSettle < 1e-3f);
-}
-
-static void testResetClearsStateButPrepareKeepsIt() {
-    VoiceFilter f;
-    f.prepare({0.5f, 0.5f, kLowPass, 0.0f}, 48000.0);
-    f.process(0, 1.0f);
-    CHECK(!f.isSilent());
-
-    // A live parameter move must not zero the state — that is what would click.
-    f.prepare({0.6f, 0.5f, kLowPass, 0.0f}, 48000.0);
-    CHECK(!f.isSilent());
-    f.prepare({0.6f, 0.5f, kBandPass, 1.0f}, 48000.0);
-    CHECK(!f.isSilent());
-
-    f.reset();
-    CHECK(f.isSilent());
-}
-
-static void testChannelStateIsIndependent() {
-    VoiceFilter f;
-    f.prepare({0.5f, 0.5f, kLowPass, 0.0f}, 48000.0);
-    f.reset();
-    f.process(0, 1.0f);
-    CHECK(f.state(0).ic2 != 0.0f);
-    CHECK(f.state(1).ic2 == 0.0f);
-
-    float frame[2] = {1.0f, -1.0f};
-    f.processFrame(frame, 2);
-    CHECK(f.state(1).ic2 < 0.0f);
-    CHECK(frame[0] != frame[1]);
-}
-
 int main() {
-    testCutoffMapsThreeDecadesLogarithmically();
-    testCutoffNormRoundTrips();
-    testQSpansPointOneToTenWithRootTwoAtCenter();
-    testQNormRoundTrips();
-    testDriveDepthIsZeroAtRestAndRisesMonotonically();
-
     testSvfCoefficientsMatchPinnedValues();
     testTheSampleRateEntersOnlyThroughG();
     testNonPositiveRatePrepareClearsStaleStateAndReportsSilent();
     testNonPositiveRatePassesSignalThroughAtEveryMorph();
 
-    testMorphEndpointMixesAreExactlyPureTaps();
-    testMorphInteriorPointMatchesCosSinNotAnAlternateEqualPowerCurve();
-    testMorphNeverBlendsHighAgainstLowPass();
     testMorphEndpointsMatchTheAnalyticTwoPoleTargets();
-    testCornerMagnitudeIsFlatAcrossTheWholeMorphSweep();
+    testCornerMagnitudeIsFlatAtQAcrossTheHighBandLowSweep();
+    testHighNotchLowCentreIsATrueNullAtTheCorner();
+    testHighNotchLowNullIsLocatedAtTheCutoff();
+    testHighNotchLowLegsHoldConstantPowerAwayFromTheNotch();
     testMorphSweepHasNoDiscontinuity();
+    testMorphLawDoesNotDisturbTheCoefficients();
+    testFilterSettingsDefaultsToTheHighBandLowLaw();
 
     testDriveZeroIsBitIdenticalToTheLinearKernel();
     testDriveZeroResponseIsLevelInvariant();
@@ -824,14 +654,6 @@ int main() {
     testResponseIsRateInvariantAtEveryMorph();
     testLowCutoffHighRateCornerHoldsTheAnalyticPeak();
 
-    testFullRangeCutoffSweepAtAudioRateStaysBounded();
-    testStateFlushesToZeroWithoutStallingInDenormals();
-    testHighPassSustainedDCDoesNotReRing();
-    testImpulseResponseMatchesTheKernel();
-    testLowpassStepSettlesToUnityAndHighpassRejectsDC();
-    testResetClearsStateButPrepareKeepsIt();
-    testChannelStateIsIndependent();
-
     if (g_fail == 0) std::printf("filter_tests: all passed\n");
     else std::printf("filter_tests: %d FAILED\n", g_fail);
     return g_fail == 0 ? 0 : 1;

tests/test_filter_morph.cpp

@@ -0,0 +1,208 @@
+// Standalone tests for the pure morph domain: normalized position -> tap weights under both
+// morph laws, and the fold of those weights into the kernel's three multipliers. Algebra only —
+// no filter is run here. Interior expectations are derived from the intended law (in radicals,
+// so they share not even a trig call with the implementation) rather than read back out of it.
+// The MEASURED consequences of each law — HP-BP-LP's flat corner, HP-notch-LP's null — live in
+// test_filter.cpp, where a filter is actually driven.
+
+#include "../src/core/instrument/engine/filter/filter_morph.h"
+#include "../src/core/instrument/engine/filter/filter_params.h"
+
+#include <cmath>
+#include <cstdio>
+#include <initializer_list>
+#include <limits>
+#include <type_traits>
+
+using namespace reasampler::instrument::engine::filter;
+
+static int g_fail = 0;
+#define CHECK(cond) do { if(!(cond)) { \
+    std::printf("FAIL line %d: %s\n", __LINE__, #cond); ++g_fail; } } while(0)
+#define CHECK_NEAR(a, b, eps) do { const double a_ = (a), b_ = (b); \
+    if (!(std::fabs(a_ - b_) <= (eps))) { \
+        std::printf("FAIL line %d: %s (%.10f) != %s (%.10f), delta %.3e\n", \
+                    __LINE__, #a, a_, #b, b_, std::fabs(a_ - b_)); ++g_fail; } } while(0)
+
+static constexpr double kPi = 3.14159265358979323846;
+
+static constexpr float kHighPass = 0.0f;
+static constexpr float kCentre = 0.5f;
+static constexpr float kLowPass = 1.0f;
+
+// cos and sin of pi/8, from the half-angle identity in radicals: cos(pi/8) = sqrt((1+cos(pi/4))/2)
+// with cos(pi/4) = sqrt(2)/2. No trig call, so nothing here is shared with filter_morph's cos/sin.
+static double cosPi8() { return std::sqrt((1.0 + std::sqrt(2.0) / 2.0) / 2.0); }  // 0.9238795325
+static double sinPi8() { return std::sqrt((1.0 - std::sqrt(2.0) / 2.0) / 2.0); }  // 0.3826834324
+
+// ---------------------------------------------------------------------------
+// Shared across both laws
+// ---------------------------------------------------------------------------
+
+// HighBandLow is enumerator 0 by design: a zero-initialized field, or one absent from an older
+// persisted blob and left default-constructed, must land on the default law rather than the SEM
+// leg. A codec written against this enum depends on that.
+static void testHighBandLowIsTheZeroEnumerator() {
+    CHECK(static_cast<std::underlying_type_t<MorphLaw>>(MorphLaw::HighBandLow) == 0);
+    CHECK(MorphLaw{} == MorphLaw::HighBandLow);
+}
+
+// The endpoints are pure taps EXACTLY under BOTH laws, not to within a rounding of cos/sin — the
+// laws differ only in the interior. Asserted on the folded mix, where "pure" is an exact
+// statement about three floats.
+static void testMorphEndpointMixesAreExactlyPureTapsUnderBothLaws() {
+    const float k = 1.0f / filterQFromNorm(0.5f);
+
+    for (MorphLaw law : {MorphLaw::HighBandLow, MorphLaw::HighNotchLow}) {
+        const MorphMix hp = morphMix(morphWeights(kHighPass, law), k);
+        CHECK(hp.m0 == 1.0f && hp.m1 == -k && hp.m2 == -1.0f);  // v0 - k*v1 - v2
+
+        const MorphMix lp = morphMix(morphWeights(kLowPass, law), k);
+        CHECK(lp.m0 == 0.0f && lp.m1 == 0.0f && lp.m2 == 1.0f);  // v2
+
+        // Out-of-range clamps to the endpoints rather than extrapolating.
+        CHECK(morphWeights(-1.0f, law).hp == 1.0f);
+        CHECK(morphWeights(2.0f, law).lp == 1.0f);
+    }
+
+    // Only the centre differs: a band-pass under one law, an HP+LP sum under the other.
+    const MorphMix bp = morphMix(morphWeights(kCentre, MorphLaw::HighBandLow), k);
+    CHECK(bp.m0 == 0.0f && bp.m1 == 1.0f && bp.m2 == 0.0f);  // v1
+
+    const MorphMix notch = morphMix(morphWeights(kCentre, MorphLaw::HighNotchLow), k);
+    CHECK(notch.m0 != 0.0f && notch.m1 != 0.0f);
+}
+
+// NaN clamps to neither endpoint (every comparison against it is false) and lands on each law's
+// degenerate — no crash, a sane fallback rather than an extrapolation. HighNotchLow has no band
+// tap to fall back to, so it lands on the leg-zero endpoint instead.
+static void testNaNFallsBackToASaneTapPerLaw() {
+    const float nan = std::numeric_limits<float>::quiet_NaN();
+
+    const MorphWeights band = morphWeights(nan, MorphLaw::HighBandLow);
+    CHECK(band.hp == 0.0f && band.bp == 1.0f && band.lp == 0.0f);
+
+    const MorphWeights sem = morphWeights(nan, MorphLaw::HighNotchLow);
+    CHECK(sem.hp == 1.0f && sem.bp == 0.0f && sem.lp == 0.0f);
+}
+
+// ---------------------------------------------------------------------------
+// HighBandLow — adjacent taps only
+// ---------------------------------------------------------------------------
+
+// Pins the cos/sin curve at an interior point, not just the endpoints and the quadrature
+// identity (hp^2+bp^2+lp^2=1, which any equal-power reparameterization would also satisfy).
+// theta=0.5*pi*t^2 (quadratic in the leg fraction, still equal-power, still exact at both
+// ends) would give hp=0.9239/bp=0.3827 here instead of the cos/sin pair's 0.7071/0.7071.
+static void testHighBandLowInteriorMatchesCosSinNotAnAlternateEqualPowerCurve() {
+    const MorphWeights w = morphWeights(0.25f, MorphLaw::HighBandLow);  // HP->BP leg, t = 0.5
+    const double theta = 0.5 * kPi * 0.5;
+    CHECK_NEAR(w.hp, std::cos(theta), 1e-6);
+    CHECK_NEAR(w.bp, std::sin(theta), 1e-6);
+    CHECK(w.lp == 0.0f);
+
+    // Each leg is half the sweep, so a leg reaches at 0.125 what the SEM's single crossfade
+    // reaches at 0.25 — the crispest algebraic statement of how the two laws differ.
+    const MorphWeights eighth = morphWeights(0.125f, MorphLaw::HighBandLow);
+    CHECK_NEAR(eighth.hp, cosPi8(), 1e-6);
+    CHECK_NEAR(eighth.bp, sinPi8(), 1e-6);
+}
+
+// HP and LP never carry weight at the same time under THIS law. That is what keeps its centre a
+// band-pass: the two are antiphase at the corner and would otherwise cancel into a notch. This
+// assertion is law-specific and is deliberately inverted for HighNotchLow below — do not relax
+// it to cover both, which would give up the guarantee entirely.
+static void testHighBandLowNeverBlendsHighAgainstLowPass() {
+    for (int i = 0; i <= 200; ++i) {
+        const MorphWeights w = morphWeights(static_cast<float>(i) / 200.0f, MorphLaw::HighBandLow);
+        CHECK(w.hp == 0.0f || w.lp == 0.0f);
+        CHECK(w.hp >= 0.0f && w.bp >= 0.0f && w.lp >= 0.0f);
+        // Equal power: the active pair sums in quadrature to unity.
+        CHECK_NEAR(w.hp * w.hp + w.bp * w.bp + w.lp * w.lp, 1.0, 1e-6);
+    }
+}
+
+// ---------------------------------------------------------------------------
+// HighNotchLow — HP against LP, which is the whole mechanism
+// ---------------------------------------------------------------------------
+
+// The exact inverse of the HighBandLow assertion above: blending HP against LP is not a defect
+// to be avoided here, it is what produces the notch. The band tap is silent throughout.
+static void testHighNotchLowBlendsHighAgainstLowPassWithNoBandTap() {
+    for (int i = 0; i <= 200; ++i) {
+        const float n = static_cast<float>(i) / 200.0f;
+        const MorphWeights w = morphWeights(n, MorphLaw::HighNotchLow);
+        CHECK(w.bp == 0.0f);
+        CHECK(w.hp >= 0.0f && w.lp >= 0.0f);
+        // Both taps carry weight everywhere strictly between the endpoints.
+        if (i > 0 && i < 200) CHECK(w.hp > 0.0f && w.lp > 0.0f);
+        // Equal power, which is what keeps the legs from sagging away from the notch frequency.
+        CHECK_NEAR(w.hp * w.hp + w.lp * w.lp, 1.0, 1e-6);
+    }
+}
+
+// Pins the single equal-power crossfade at interior points against radical-derived values, so a
+// law that still hits both endpoints but bends differently between them fails. Discriminators at
+// n=0.25: a LINEAR crossfade gives 0.75/0.25; a two-leg construction (HighBandLow's spacing
+// applied to an HP/LP pair) gives 0.7071/0.7071. Both are far outside this tolerance.
+static void testHighNotchLowInteriorWeightsMatchTheSingleEqualPowerCrossfade() {
+    // cos/sin of pi/8 and 3pi/8; the latter pair is the former swapped.
+    const double c8 = cosPi8(), s8 = sinPi8();
+    CHECK_NEAR(c8, 0.9238795325112867, 1e-15);
+    CHECK_NEAR(s8, 0.3826834323650898, 1e-15);
+
+    const MorphWeights quarter = morphWeights(0.25f, MorphLaw::HighNotchLow);
+    CHECK_NEAR(quarter.hp, c8, 1e-6);
+    CHECK_NEAR(quarter.lp, s8, 1e-6);
+
+    const MorphWeights threeQuarters = morphWeights(0.75f, MorphLaw::HighNotchLow);
+    CHECK_NEAR(threeQuarters.hp, s8, 1e-6);
+    CHECK_NEAR(threeQuarters.lp, c8, 1e-6);
+
+    const MorphWeights centre = morphWeights(kCentre, MorphLaw::HighNotchLow);
+    CHECK_NEAR(centre.hp, std::sqrt(2.0) / 2.0, 1e-6);
+    CHECK_NEAR(centre.lp, std::sqrt(2.0) / 2.0, 1e-6);
+
+    // Symmetric about the centre, so the sweep reads the same in either direction.
+    for (int i = 0; i <= 100; ++i) {
+        const float n = static_cast<float>(i) / 100.0f;
+        const MorphWeights a = morphWeights(n, MorphLaw::HighNotchLow);
+        const MorphWeights b = morphWeights(1.0f - n, MorphLaw::HighNotchLow);
+        CHECK_NEAR(a.hp, b.lp, 1e-6);
+    }
+}
+
+// The centre's cancellation is STRUCTURAL, not a runtime near-miss of two large numbers. The fold
+// is m2 = lp - hp, and at the centre the two weights are the same float — cos and sin of pi/4
+// differ by about an ulp of DOUBLE, ~1e-16, which is nine orders below float's ~6e-8 spacing
+// there, so they round to one value. m2 is therefore exactly 0 and the output reduces to
+// hp*(x - k*v1): the high and low taps cannot drift apart by a rounding.
+static void testHighNotchLowCentreFoldsToAnExactlyCancellingMix() {
+    for (float res : {0.0f, 0.5f, 1.0f}) {
+        const float k = 1.0f / filterQFromNorm(res);
+        const MorphWeights w = morphWeights(kCentre, MorphLaw::HighNotchLow);
+        CHECK(w.hp == w.lp);
+
+        const MorphMix m = morphMix(w, k);
+        CHECK(m.m2 == 0.0f);
+        CHECK(m.m0 == w.hp);
+        CHECK(m.m1 == -w.hp * k);
+    }
+}
+
+int main() {
+    testHighBandLowIsTheZeroEnumerator();
+    testMorphEndpointMixesAreExactlyPureTapsUnderBothLaws();
+    testNaNFallsBackToASaneTapPerLaw();
+
+    testHighBandLowInteriorMatchesCosSinNotAnAlternateEqualPowerCurve();
+    testHighBandLowNeverBlendsHighAgainstLowPass();
+
+    testHighNotchLowBlendsHighAgainstLowPassWithNoBandTap();
+    testHighNotchLowInteriorWeightsMatchTheSingleEqualPowerCrossfade();
+    testHighNotchLowCentreFoldsToAnExactlyCancellingMix();
+
+    if (g_fail == 0) std::printf("filter_morph_tests: all passed\n");
+    else std::printf("filter_morph_tests: %d FAILED\n", g_fail);
+    return g_fail == 0 ? 0 : 1;
+}

tests/test_filter_params.cpp

@@ -0,0 +1,121 @@
+// Standalone tests for the filter's control domain — normalized knob position to cutoff Hz, Q,
+// and drive depth, plus the exact inverses. No DSP is run here and no sample rate appears, which
+// is the point: filter_params is deliberately rate-free. Interior points are pinned against a
+// derivation of the intended law written out in-test, so a curve that still hits the anchors but
+// bends differently between them fails.
+
+#include "../src/core/instrument/engine/filter/filter_params.h"
+
+#include <cmath>
+#include <cstdio>
+
+using namespace reasampler::instrument::engine::filter;
+
+static int g_fail = 0;
+#define CHECK(cond) do { if(!(cond)) { \
+    std::printf("FAIL line %d: %s\n", __LINE__, #cond); ++g_fail; } } while(0)
+#define CHECK_NEAR(a, b, eps) do { const double a_ = (a), b_ = (b); \
+    if (!(std::fabs(a_ - b_) <= (eps))) { \
+        std::printf("FAIL line %d: %s (%.10f) != %s (%.10f), delta %.3e\n", \
+                    __LINE__, #a, a_, #b, b_, std::fabs(a_ - b_)); ++g_fail; } } while(0)
+
+static void testCutoffMapsThreeDecadesLogarithmically() {
+    CHECK_NEAR(filterCutoffHzFromNorm(0.0f), 20.0, 1e-3);
+    CHECK_NEAR(filterCutoffHzFromNorm(1.0f), 20000.0, 1e-2);
+
+    // Exactly three decades, so the decade midpoints land on round numbers.
+    CHECK_NEAR(filterCutoffHzFromNorm(1.0f / 3.0f), 200.0, 1e-3);
+    CHECK_NEAR(filterCutoffHzFromNorm(2.0f / 3.0f), 2000.0, 1e-2);
+
+    // Half-decade steps confirm the sweep is log, not linear.
+    CHECK_NEAR(filterCutoffHzFromNorm(1.0f / 6.0f), 20.0 * std::sqrt(10.0), 1e-3);
+    CHECK_NEAR(filterCutoffHzFromNorm(0.5f), 20.0 * std::sqrt(1000.0), 1e-2);
+
+    CHECK_NEAR(filterCutoffHzFromNorm(-1.0f), 20.0, 1e-3);
+    CHECK_NEAR(filterCutoffHzFromNorm(2.0f), 20000.0, 1e-2);
+}
+
+static void testCutoffNormRoundTrips() {
+    for (int i = 0; i <= 20; ++i) {
+        const float n = static_cast<float>(i) / 20.0f;
+        CHECK_NEAR(filterNormFromCutoffHz(filterCutoffHzFromNorm(n)), n, 1e-6);
+    }
+    CHECK_NEAR(filterNormFromCutoffHz(200.0f), 1.0 / 3.0, 1e-6);
+    CHECK(filterNormFromCutoffHz(1.0f) == 0.0f);
+    CHECK(filterNormFromCutoffHz(48000.0f) == 1.0f);
+}
+
+static void testQSpansPointOneToTenWithRootTwoAtCenter() {
+    CHECK_NEAR(filterQFromNorm(0.0f), 0.1, 1e-6);
+    CHECK_NEAR(filterQFromNorm(0.5f), std::sqrt(2.0), 1e-5);
+    CHECK_NEAR(filterQFromNorm(1.0f), 10.0, 1e-4);
+
+    CHECK_NEAR(filterQFromNorm(-1.0f), 0.1, 1e-6);
+    CHECK_NEAR(filterQFromNorm(2.0f), 10.0, 1e-4);
+
+    // Pins the single quadratic-in-log-Q curve at two interior points, derived independently by
+    // solving log Q = a + b*n + c*n^2 through the three anchors above rather than read out of
+    // the implementation. A two-spliced-log-segments curve (log-linear on each half, the design
+    // this module doc explicitly rejects for its center-detent slope kink) would give 0.376 and
+    // 3.761 here instead — both comfortably outside this tolerance.
+    {
+        const double lo = std::log(static_cast<double>(kFilterQMin));
+        const double mid = std::log(static_cast<double>(kFilterQCenter));
+        const double hi = std::log(static_cast<double>(kFilterQMax));
+        const double c = 2.0 * lo + 2.0 * hi - 4.0 * mid;
+        const double b = hi - lo - c;
+        const double a = lo;
+        auto qLaw = [&](double n) { return std::exp(a + b * n + c * n * n); };
+        CHECK_NEAR(filterQFromNorm(0.25f), qLaw(0.25), 1e-5);
+        CHECK_NEAR(filterQFromNorm(0.75f), qLaw(0.75), 1e-5);
+    }
+
+    // Strictly monotonic across the whole travel — no fold-back from the quadratic term.
+    float prev = -1.0f;
+    for (int i = 0; i <= 1000; ++i) {
+        const float q = filterQFromNorm(static_cast<float>(i) / 1000.0f);
+        CHECK(q > prev);
+        prev = q;
+    }
+}
+
+static void testQNormRoundTrips() {
+    for (int i = 0; i <= 20; ++i) {
+        const float n = static_cast<float>(i) / 20.0f;
+        CHECK_NEAR(filterNormFromQ(filterQFromNorm(n)), n, 1e-5);
+    }
+    CHECK_NEAR(filterNormFromQ(static_cast<float>(std::sqrt(2.0))), 0.5, 1e-5);
+    CHECK(filterNormFromQ(0.0f) == 0.0f);
+    CHECK(filterNormFromQ(1000.0f) == 1.0f);
+}
+
+static void testDriveDepthIsZeroAtRestAndRisesMonotonically() {
+    // Exactly zero, not nearly: the limiter is the identity only at depth 0.
+    CHECK(filterDriveDepthFromNorm(0.0f) == 0.0f);
+    CHECK(filterDriveDepthFromNorm(-1.0f) == 0.0f);
+    CHECK_NEAR(filterDriveDepthFromNorm(1.0f), kFilterDriveDepthMax, 1e-6);
+    CHECK_NEAR(filterDriveDepthFromNorm(2.0f), kFilterDriveDepthMax, 1e-6);
+
+    // Pins the SQUARE law at an interior point, not just the anchors: a linear law would give
+    // kFilterDriveDepthMax/2 (2.0) here, not kFilterDriveDepthMax/4 (1.0).
+    CHECK_NEAR(filterDriveDepthFromNorm(0.5f), kFilterDriveDepthMax * 0.25, 1e-6);
+
+    float prev = -1.0f;
+    for (int i = 0; i <= 100; ++i) {
+        const float d = filterDriveDepthFromNorm(static_cast<float>(i) / 100.0f);
+        CHECK(d > prev);
+        prev = d;
+    }
+}
+
+int main() {
+    testCutoffMapsThreeDecadesLogarithmically();
+    testCutoffNormRoundTrips();
+    testQSpansPointOneToTenWithRootTwoAtCenter();
+    testQNormRoundTrips();
+    testDriveDepthIsZeroAtRestAndRisesMonotonically();
+
+    if (g_fail == 0) std::printf("filter_params_tests: all passed\n");
+    else std::printf("filter_params_tests: %d FAILED\n", g_fail);
+    return g_fail == 0 ? 0 : 1;
+}

tests/test_filter_state.cpp

@@ -0,0 +1,247 @@
+// Standalone tests for the running filter's NUMERICAL behaviour and state lifecycle — bounded
+// output under a live parameter sweep, the denormal flush, DC handling, the impulse response
+// against the coefficients, and reset/prepare/per-channel state rules. Split from test_filter.cpp
+// along the one seam that costs nothing: none of these need the frequency-response measurement
+// harness, so the analytic reference lives in exactly one file and cannot fork.
+
+#include "../src/core/instrument/engine/filter/filter_coeffs.h"
+#include "../src/core/instrument/engine/filter/filter_morph.h"
+#include "../src/core/instrument/engine/filter/filter_params.h"
+#include "../src/core/instrument/engine/filter/voice_filter.h"
+
+#include <cfloat>
+#include <cmath>
+#include <cstdio>
+#include <initializer_list>
+
+using namespace reasampler::instrument::engine::filter;
+
+static int g_fail = 0;
+#define CHECK(cond) do { if(!(cond)) { \
+    std::printf("FAIL line %d: %s\n", __LINE__, #cond); ++g_fail; } } while(0)
+#define CHECK_NEAR(a, b, eps) do { const double a_ = (a), b_ = (b); \
+    if (!(std::fabs(a_ - b_) <= (eps))) { \
+        std::printf("FAIL line %d: %s (%.10f) != %s (%.10f), delta %.3e\n", \
+                    __LINE__, #a, a_, #b, b_, std::fabs(a_ - b_)); ++g_fail; } } while(0)
+
+static constexpr double kPi = 3.14159265358979323846;
+
+static constexpr float kHighPass = 0.0f;
+static constexpr float kBandPass = 0.5f;
+static constexpr float kCentre = 0.5f;
+static constexpr float kLowPass = 1.0f;
+
+static const MorphLaw kBothLaws[] = {MorphLaw::HighBandLow, MorphLaw::HighNotchLow};
+
+// The rates the invariance claims are made over.
+static const double kRates[] = {44100.0, 48000.0, 88200.0, 96000.0, 192000.0};
+static constexpr int kRateCount = 5;
+
+static FilterSettings at(double fcHz, float res, float morph, float drive = 0.0f,
+                         MorphLaw law = MorphLaw::HighBandLow) {
+    return {filterNormFromCutoffHz(static_cast<float>(fcHz)), res, morph, drive, law};
+}
+
+static void testFullRangeCutoffSweepAtAudioRateStaysBounded() {
+    unsigned rng = 0x13579bdfu;
+    auto noise = [&rng]() {
+        rng = rng * 1664525u + 1013904223u;
+        return static_cast<float>(static_cast<int>(rng >> 9) - (1 << 22)) /
+               static_cast<float>(1 << 22);
+    };
+
+    for (int r = 0; r < kRateCount; ++r) {
+        const double sr = kRates[r];
+        for (MorphLaw law : kBothLaws) {
+            for (float morph : {kHighPass, kBandPass, kLowPass}) {
+                for (float res : {0.0f, 1.0f}) {
+                    for (float drive : {0.0f, 1.0f}) {
+                        VoiceFilter f;
+                        f.reset();
+                        // A fixed WALL-CLOCK sweep: the same cutoff travel per second at every
+                        // rate, so the per-sample coefficient step gets no gentler as the rate
+                        // rises.
+                        const int n = static_cast<int>(sr * 0.25);
+                        for (int i = 0; i < n; ++i) {
+                            const float t = static_cast<float>(i) / static_cast<float>(n - 1);
+                            f.prepare({t, res, morph, drive, law}, sr);
+                            const float y = f.process(0, noise());
+                            CHECK(std::isfinite(y));
+                            CHECK(std::fabs(y) < 100.0f);
+                            if (!std::isfinite(y)) return;  // stop before the log floods
+                        }
+                    }
+                }
+            }
+        }
+    }
+}
+
+// The flush tests the ENVELOPE — both integrators — not one sample. ic1 and ic2 are in
+// quadrature, so a resonator swings each through zero twice a cycle; flushing on a single one
+// injects a step in phase with the resonance, which the resonance amplifies, and the filter
+// limit-cycles at the floor forever instead of going quiet. Re-verified for TPT rather than
+// assumed to carry over from the retired Direct Form I state.
+static void checkFlushGoesSilent(double sr, float morph, float drive, MorphLaw law) {
+    // The decay to the floor is a fixed WALL-CLOCK time, so the budget scales with the rate.
+    const int budget = static_cast<int>(sr * 0.5);
+    VoiceFilter f;
+    f.prepare(at(1000.0, 1.0f, morph, drive, law), sr);
+    f.reset();
+
+    // Excite, then hard-cut to silence the way a released voice does.
+    const int excite = static_cast<int>(sr * 0.01);
+    for (int i = 0; i < excite; ++i) {
+        f.process(0, 0.5f * static_cast<float>(std::sin(2.0 * kPi * 1000.0 * i / sr)));
+    }
+
+    int subnormalSamples = 0, silentAt = -1;
+    for (int i = 0; i < budget; ++i) {
+        f.process(0, 0.0f);
+        const VoiceFilter::State& s = f.state(0);
+        if ((s.ic1 != 0.0f && std::fabs(s.ic1) < FLT_MIN) ||
+            (s.ic2 != 0.0f && std::fabs(s.ic2) < FLT_MIN)) {
+            ++subnormalSamples;
+        }
+        if (silentAt < 0 && f.isSilent()) silentAt = i;
+    }
+    // Without the flush the state grinds down through the subnormal range for thousands of
+    // samples; a stray sample or two at a zero crossing is not a stall.
+    CHECK(subnormalSamples <= 2);
+    CHECK(silentAt >= 0);
+    CHECK(silentAt < budget);
+    // And it stays silent — a flush that perturbs the loop would re-excite it.
+    for (int i = 0; i < 1000; ++i) CHECK(f.process(0, 0.0f) == 0.0f);
+    CHECK(f.isSilent());
+}
+
+static void testStateFlushesToZeroWithoutStallingInDenormals() {
+    for (int r = 0; r < kRateCount; ++r) {
+        for (MorphLaw law : kBothLaws) {
+            for (float morph : {kHighPass, kBandPass, kLowPass}) {
+                for (float drive : {0.0f, 1.0f}) checkFlushGoesSilent(kRates[r], morph, drive, law);
+            }
+        }
+    }
+}
+
+// A high-pass under sustained DC must settle to zero and STAY there. Sampling only the final
+// value is not enough: a resonator swings through zero twice a cycle, so a single late sample
+// can land near zero while the envelope still rings well above it. This regressed a click train
+// on the retired topology, where flushing the FIR history discarded the pinned DC and the next
+// sample recomputed a full-amplitude step. TPT has no FIR history to discard, so the hazard is
+// structural rather than a tuning — but the assertion is cheap and pins the outcome. Morph 0 is
+// the same pure high-pass under either law, so this needs no law loop.
+static void testHighPassSustainedDCDoesNotReRing() {
+    for (int r = 0; r < kRateCount; ++r) {
+        const double sr = kRates[r];
+        for (float drive : {0.0f, 1.0f}) {
+            VoiceFilter f;
+            f.prepare(at(1000.0, 1.0f, kHighPass, drive), sr);
+            f.reset();
+            const int settle = static_cast<int>(sr * 0.05);
+            float worstAfterSettle = 0.0f;
+            for (int i = 0; i < static_cast<int>(sr * 0.5); ++i) {
+                const float y = f.process(0, 1.0f);
+                if (i >= settle) worstAfterSettle = std::fmax(worstAfterSettle, std::fabs(y));
+            }
+            CHECK(worstAfterSettle < 1e-3f);
+        }
+    }
+}
+
+static void testImpulseResponseMatchesTheKernel() {
+    VoiceFilter f;
+    f.prepare(at(1000.0, 0.5f, kLowPass), 48000.0);
+    f.reset();
+    const SvfCoeffs c = f.coeffs();
+    // From a cleared state the first sample reduces to the coefficients alone: v1 == a2, v2 == a3.
+    CHECK_NEAR(f.process(0, 1.0f), c.a3, 1e-7);
+
+    VoiceFilter bp;
+    bp.prepare(at(1000.0, 0.5f, kBandPass), 48000.0);
+    bp.reset();
+    CHECK_NEAR(bp.process(0, 1.0f), c.a2, 1e-7);
+
+    VoiceFilter hp;
+    hp.prepare(at(1000.0, 0.5f, kHighPass), 48000.0);
+    hp.reset();
+    CHECK_NEAR(hp.process(0, 1.0f), 1.0 - c.k * c.a2 - c.a3, 1e-7);
+
+    // Under HighNotchLow the centre's first sample is the SUM of the high and low taps, scaled by
+    // the shared weight — the same algebra the null rests on, seen one sample in.
+    VoiceFilter sem;
+    sem.prepare(at(1000.0, 0.5f, kCentre, 0.0f, MorphLaw::HighNotchLow), 48000.0);
+    sem.reset();
+    const double w = morphWeights(kCentre, MorphLaw::HighNotchLow).hp;
+    const double highTap = 1.0 - c.k * c.a2 - c.a3;
+    const double lowTap = c.a3;
+    CHECK_NEAR(sem.process(0, 1.0f), w * (highTap + lowTap), 1e-6);
+}
+
+static void testLowpassStepSettlesToUnityAndHighpassRejectsDC() {
+    const double sr = 48000.0;
+    VoiceFilter f;
+    f.prepare(at(1000.0, 0.0f, kLowPass), sr);
+    f.reset();
+    float y = 0.0f;
+    for (int i = 0; i < 48000; ++i) y = f.process(0, 1.0f);
+    CHECK_NEAR(y, 1.0, 1e-3);  // DC passes a lowpass at unity
+
+    VoiceFilter hp;
+    hp.prepare(at(1000.0, 0.0f, kHighPass), sr);
+    hp.reset();
+    float worstAfterSettle = 0.0f;
+    for (int i = 0; i < 48000; ++i) {
+        y = hp.process(0, 1.0f);
+        if (i >= 200) worstAfterSettle = std::fmax(worstAfterSettle, std::fabs(y));
+    }
+    CHECK(worstAfterSettle < 1e-3f);
+}
+
+static void testResetClearsStateButPrepareKeepsIt() {
+    VoiceFilter f;
+    f.prepare({0.5f, 0.5f, kLowPass, 0.0f}, 48000.0);
+    f.process(0, 1.0f);
+    CHECK(!f.isSilent());
+
+    // A live parameter move must not zero the state — that is what would click. Switching the
+    // morph law is a parameter move like any other: it only recomputes the mix.
+    f.prepare({0.6f, 0.5f, kLowPass, 0.0f}, 48000.0);
+    CHECK(!f.isSilent());
+    f.prepare({0.6f, 0.5f, kBandPass, 1.0f}, 48000.0);
+    CHECK(!f.isSilent());
+    f.prepare({0.6f, 0.5f, kCentre, 1.0f, MorphLaw::HighNotchLow}, 48000.0);
+    CHECK(!f.isSilent());
+
+    f.reset();
+    CHECK(f.isSilent());
+}
+
+static void testChannelStateIsIndependent() {
+    VoiceFilter f;
+    f.prepare({0.5f, 0.5f, kLowPass, 0.0f}, 48000.0);
+    f.reset();
+    f.process(0, 1.0f);
+    CHECK(f.state(0).ic2 != 0.0f);
+    CHECK(f.state(1).ic2 == 0.0f);
+
+    float frame[2] = {1.0f, -1.0f};
+    f.processFrame(frame, 2);
+    CHECK(f.state(1).ic2 < 0.0f);
+    CHECK(frame[0] != frame[1]);
+}
+
+int main() {
+    testFullRangeCutoffSweepAtAudioRateStaysBounded();
+    testStateFlushesToZeroWithoutStallingInDenormals();
+    testHighPassSustainedDCDoesNotReRing();
+    testImpulseResponseMatchesTheKernel();
+    testLowpassStepSettlesToUnityAndHighpassRejectsDC();
+    testResetClearsStateButPrepareKeepsIt();
+    testChannelStateIsIndependent();
+
+    if (g_fail == 0) std::printf("filter_state_tests: all passed\n");
+    else std::printf("filter_state_tests: %d FAILED\n", g_fail);
+    return g_fail == 0 ? 0 : 1;
+}