CODEX: Validate resident pose LMA decisions

Co-authored-by: Codex <codex@elphel.com>

src/main/java/com/elphel/imagej/cuas/rt/CuasPoseRT.java

@@ -1035,15 +1035,16 @@ public class CuasPoseRT {
 			return result;
 		}); // first production gate: replace only getFxDerivs per-tile work; callers remain unchanged // By Codex on 07/15/2026
 		intersceneLma.setPoseLmaResidentStepProvider((
-				lambda, weights, y, eigen, vector, pureWeight, numTiles, capturePrepared) -> {
+				lambda, weights, y, eigen, vector, pureWeight, rmsDiff, numTiles, capturePrepared) -> {
 			final IntersceneLmaFloat.ResidentLmaStepResult result =
 					lmaGpu.execPoseLmaResidentStep(
-							lambda, weights, y, eigen, vector, pureWeight,
+							lambda, weights, y, eigen, vector, pureWeight, rmsDiff,
 							numTiles, capturePrepared);
 			if ((result != null) && !pose_lma_step_path_reported) {
 				pose_lma_step_path_reported = true;
-				System.out.println("CuasPoseRT: resident CUDA float LMA candidate/RMS oracle active "+
-						"(separate candidate fx/J buffers; Java acceptance authoritative)");
+				System.out.println("CuasPoseRT: resident CUDA float LMA decision/state oracle active "+
+						"(GPU candidate accept swaps raw buffers; reject retains current; "+
+						"Java acceptance authoritative)");
 			}
 			return result;
 		}); // production candidate: resident raw fx/J -> parallel prepare/reduce -> fixed 3x3 tail // By Codex on 07/15/2026

src/main/java/com/elphel/imagej/gpu/GpuQuad.java

@@ -2949,6 +2949,7 @@ public class GpuQuad{ // quad camera description
 			float [] eigen,
 			float [] vector,
 			float pure_weight,
+			float rms_diff,
 			int num_tiles,
 			boolean capture_prepared) {
 		return null;

src/main/java/com/elphel/imagej/gpu/jna/GpuQuadJna.java

@@ -564,6 +564,7 @@ public class GpuQuadJna extends GpuQuad {
             final float[] eigen,
             final float[] vector,
             final float pureWeight,
+            final float rmsDiff,
             final int numTiles,
             final boolean capturePrepared) {
         final int numValues = IntersceneLmaFloat.NUM_COMPONENTS * numTiles +
@@ -572,7 +573,8 @@ public class GpuQuadJna extends GpuQuad {
                 (y == null) || (y.length != numValues) || (eigen == null) ||
                 (eigen.length != 4 * numTiles) || (vector == null) ||
                 (vector.length != IntersceneLmaFloat.NUM_PARAMS) ||
-                !Float.isFinite(pureWeight) || !(pureWeight > 0.0f)) {
+                !Float.isFinite(pureWeight) || !(pureWeight > 0.0f) ||
+                !Float.isFinite(rmsDiff)) {
             throw new IllegalArgumentException("GpuQuadJna.execPoseLmaResidentStep: inconsistent inputs");
         }
         final float[] packed = new float[IntersceneLmaFloat.LMA_RESIDENT_RESULT_FLOATS];
@@ -580,7 +582,7 @@ public class GpuQuadJna extends GpuQuad {
                 new float[IntersceneLmaFloat.NUM_PARAMS * numValues] : null;
         final float[] preparedYmfx = capturePrepared ? new float[numValues] : null;
         final int rc = lib.tp_proc_exec_pose_lma_resident_step(
-                proc, lambda, weights, y, eigen, vector, pureWeight, numTiles,
+                proc, lambda, weights, y, eigen, vector, pureWeight, rmsDiff, numTiles,
                 capturePrepared ? 1 : 0, packed, preparedJt, preparedYmfx);
         if (rc != 0) {
             throw new IllegalStateException("GpuQuadJna.execPoseLmaResidentStep rc="+rc+": "+
@@ -596,7 +598,9 @@ public class GpuQuadJna extends GpuQuad {
                 packed[IntersceneLmaFloat.LMA_RESULT_FLOATS],
                 packed[IntersceneLmaFloat.LMA_RESULT_FLOATS + 1],
                 packed[IntersceneLmaFloat.LMA_RESULT_FLOATS + 2],
-                packed[IntersceneLmaFloat.LMA_RESULT_FLOATS + 3]);
+                packed[IntersceneLmaFloat.LMA_RESULT_FLOATS + 3],
+                packed[IntersceneLmaFloat.LMA_RESULT_FLOATS + 4] != 0.0f,
+                packed[IntersceneLmaFloat.LMA_RESULT_FLOATS + 5] != 0.0f);
     }
 
     private static IntersceneLmaFloat.LmaStepResult unpackPoseLmaStep(final float[] packed) {

src/main/java/com/elphel/imagej/gpu/jna/TpJna.java

@@ -155,12 +155,12 @@ public interface TpJna extends Library {
                                       float[] ymfxWeighted, float[] vector,
                                       int numValues, float[] result);
     /** Continue from raw fx/J left by the most recent pose kernel. Result has
-     *  the 19 candidate floats followed by current and candidate full/pure RMS. Prepared
-     *  arrays are nullable unless capturePrepared is nonzero. */
+     *  the 19 candidate floats, current/candidate full/pure RMS, then accepted/stop.
+     *  Prepared arrays are nullable unless capturePrepared is nonzero. */
     int     tp_proc_exec_pose_lma_resident_step(Pointer proc,
                                       float lambda, float[] weights, float[] y,
                                       float[] eigen, float[] vector, float pureWeight,
-                                      int numTiles, int capturePrepared, float[] result,
+                                      float rmsDiff, int numTiles, int capturePrepared, float[] result,
                                       float[] preparedJt, float[] preparedYmfx);
     /** Deterministic double normal-equation products for the lean pose LMA.
      *  jt is parameter-major [numParams][numValues]; out is row-major H followed by b.

src/main/java/com/elphel/imagej/tileprocessor/IntersceneLma.java

@@ -121,12 +121,14 @@ public class IntersceneLma {
 				float [] eigen,
 				float [] vector,
 				float pureWeight,
+				float rmsDiff,
 				int numTiles,
 				boolean capturePrepared);
 	}
 	private PoseLmaResidentStepProvider poseLmaResidentStepProvider = null;
 	private static boolean poseLmaStepOracleReported = false;
 	private static boolean poseLmaAcceptanceOracleReported = false;
+	private static boolean poseLmaAcceptanceMismatchReported = false;
 	private IntersceneLmaFloat.Camera poseFloatReference = null;
 	private IntersceneLmaFloat.Camera poseFloatScene = null;
 	private float [] poseFloatCenters = null;
@@ -944,7 +946,8 @@ public class IntersceneLma {
 					eig_trans, poseFloatLastResult.numTiles);
 			residentResult = poseLmaResidentStepProvider.calculate(
 					(float) lambda, floatWeights, floatY, floatEigen, floatVector,
-					(float) pure_weight, poseFloatLastResult.numTiles, captureResidentOracle);
+					(float) pure_weight, (float) rms_diff, poseFloatLastResult.numTiles,
+					captureResidentOracle);
 			if (residentResult != null) residentPoseStep = residentResult.step;
 			if (captureResidentOracle && (residentPoseStep != null)) {
 				if (residentResult.prepared == null) {
@@ -1047,9 +1050,16 @@ public class IntersceneLma {
 		if (last_ymfx == null) {
 			return null; // need to re-init/restart LMA
 		}
-		if (residentCandidateValid && (residentResult != null) && (debug_level >= 1) &&
-				!poseLmaAcceptanceOracleReported) {
-			reportPoseRmsDecision(residentResult, this.last_rms, rms, rms_diff);
+		if (residentCandidateValid && (residentResult != null) && (debug_level >= 1)) {
+			final boolean decisionsMatch = poseRmsDecisionsMatch(
+					residentResult, this.last_rms, rms, rms_diff);
+			if (!poseLmaAcceptanceOracleReported ||
+					(!decisionsMatch && !poseLmaAcceptanceMismatchReported)) {
+				reportPoseRmsDecision(
+						decisionsMatch ? "" : " MISMATCH",
+						residentResult, this.last_rms, rms, rms_diff);
+			}
+			if (!decisionsMatch) poseLmaAcceptanceMismatchReported = true;
 			poseLmaAcceptanceOracleReported = true;
 		}
 
@@ -1934,7 +1944,23 @@ public class IntersceneLma {
 				label, poseStep.valid, maxProducts, maxDelta, maxCandidate));
 	}
 
+	private static boolean poseRmsDecisionsMatch(
+			final IntersceneLmaFloat.ResidentLmaStepResult resident,
+			final double [] javaCurrent,
+			final double [] javaCandidate,
+			final double rmsDiff) {
+		final boolean doubleImproved = javaCandidate[0] < javaCurrent[0];
+		final boolean doubleStop = doubleImproved &&
+				(javaCandidate[0] >= javaCurrent[0] * (1.0 - rmsDiff));
+		return Float.isFinite(resident.currentRms) &&
+				Float.isFinite(resident.currentPureRms) &&
+				Float.isFinite(resident.candidateRms) &&
+				Float.isFinite(resident.candidatePureRms) &&
+				(resident.accepted == doubleImproved) && (resident.stop == doubleStop);
+	}
+
 	private static void reportPoseRmsDecision(
+			final String suffix,
 			final IntersceneLmaFloat.ResidentLmaStepResult resident,
 			final double [] javaCurrent,
 			final double [] javaCandidate,
@@ -1943,21 +1969,19 @@ public class IntersceneLma {
 				Float.isFinite(resident.currentPureRms) &&
 				Float.isFinite(resident.candidateRms) &&
 				Float.isFinite(resident.candidatePureRms);
-		final boolean floatImproved = finite && (resident.candidateRms < resident.currentRms);
 		final boolean doubleImproved = javaCandidate[0] < javaCurrent[0];
-		final boolean floatStop = floatImproved &&
-				(resident.candidateRms >= resident.currentRms * (1.0f - (float) rmsDiff));
 		final boolean doubleStop = doubleImproved &&
 				(javaCandidate[0] >= javaCurrent[0] * (1.0 - rmsDiff));
 		System.out.println(String.format(
-				"IntersceneLma resident CUDA-float vs Java-double RMS/decision: finite=%s, |current RMS|=%g, |current pure RMS|=%g, |candidate RMS|=%g, |candidate pure RMS|=%g, improved-match=%s, stop-match=%s",
+				"IntersceneLma resident CUDA-float vs Java-double RMS/decision%s: finite=%s, |current RMS|=%g, |current pure RMS|=%g, |candidate RMS|=%g, |candidate pure RMS|=%g, improved-match=%s, stop-match=%s",
+				suffix,
 				finite,
 				Math.abs(resident.currentRms - javaCurrent[0]),
 				Math.abs(resident.currentPureRms - javaCurrent[1]),
 				Math.abs(resident.candidateRms - javaCandidate[0]),
 				Math.abs(resident.candidatePureRms - javaCandidate[1]),
-				floatImproved == doubleImproved,
-				floatStop == doubleStop));
+				resident.accepted == doubleImproved,
+				resident.stop == doubleStop));
 	}
 
 	private static final class DoubleLmaStepResult {

src/main/java/com/elphel/imagej/tileprocessor/IntersceneLmaFloat.java

@@ -42,7 +42,7 @@ public final class IntersceneLmaFloat {
 	public static final int NUM_COMPONENTS = 2;
 	public static final int LMA_PRODUCTS = NUM_PARAMS * NUM_PARAMS + NUM_PARAMS;
 	public static final int LMA_RESULT_FLOATS = 1 + LMA_PRODUCTS + 2 * NUM_PARAMS;
-	public static final int LMA_RESIDENT_RESULT_FLOATS = LMA_RESULT_FLOATS + 4;
+	public static final int LMA_RESIDENT_RESULT_FLOATS = LMA_RESULT_FLOATS + 6;
 	public static final float INFINITY_DISPARITY = 0.01f;
 
 	static final int ERS_XYZ = 0;
@@ -220,6 +220,8 @@ public final class IntersceneLmaFloat {
 		public final float currentPureRms;
 		public final float candidateRms;
 		public final float candidatePureRms;
+		public final boolean accepted;
+		public final boolean stop;
 
 		public ResidentLmaStepResult(
 				final LmaStepResult step,
@@ -227,13 +229,17 @@ public final class IntersceneLmaFloat {
 				final float currentRms,
 				final float currentPureRms,
 				final float candidateRms,
-				final float candidatePureRms) {
+				final float candidatePureRms,
+				final boolean accepted,
+				final boolean stop) {
 			this.step = step;
 			this.prepared = prepared;
 			this.currentRms = currentRms;
 			this.currentPureRms = currentPureRms;
 			this.candidateRms = candidateRms;
 			this.candidatePureRms = candidatePureRms;
+			this.accepted = accepted;
+			this.stop = stop;
 		}
 	}