CLAUDE: MB convert erase-first overload - undefined tiles reach consolidation as NaN

Per Andrey: undefined tiles (= not in the task list; the fixed-length 5120
arrays represent them as nulls) must be handled by the channel consolidation
- weighted-by-existing average, all-NaN tile when no sensor has it (now Java
CuasTD.consolidateSensorsTD, later the CUDA clt_average_sensors kernel).
The consolidation detects missing by NaN, but the MB double convert ran with
erase=-1, leaving STALE (valid-looking) data at undefined tiles.

- ImageDtt.interCorrTDMotionBlur: new erase-capable overload (erase_clt_first:
  >0 NaN, 0 zeros, <0 legacy no-erase) - erases before the FIRST (SET) convert
  only; the SECOND (negative-scale SUBTRACT) never erases. Legacy signature
  delegates with -1: oracle/SfM callers byte-identical.
- CuasRender MB path uses erase_clt_first=1: the GPU TD buffer's NaN pattern
  now mirrors the task-list membership, so consolidation + renders see
  undefined tiles as NaN - never the previous scene.

Verified: mvn -DskipTests clean package OK.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

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

@@ -115,9 +115,9 @@ public class CuasRender {
 		if (mb_vectors != null) {
 			// Motion-blur compensation: TWO task sets (original positive scale + shifted
 			// negative scale), double convert_direct - the oracle machinery
-			// (setInterTasksMotionBlur / interCorrTDMotionBlur). NOTE: no erased re-convert
-			// here (it would destroy the two-pass accumulation) - tiles outside the task
-			// set keep the previous scene's data; comparisons must mask to task tiles.
+			// (setInterTasksMotionBlur / interCorrTDMotionBlur). The FIRST convert erases
+			// the CLT to NaN, so tiles absent from the task set reach the NaN-aware channel
+			// consolidation (and the renders) as NaN - never stale data (Andrey 07/05/2026).
 			final TpTask [][] tp_tasks2 = GpuQuad.setInterTasksMotionBlur(
 					scene.getNumSensors(),
 					img_width,
@@ -138,6 +138,7 @@ public class CuasRender {
 			}
 			image_dtt.interCorrTDMotionBlur(
 					clt_parameters.img_dtt,
+					1,                          // erase_clt_first: NaN outside the task set // By Claude on 07/05/2026
 					tp_tasks2,
 					null,
 					clt_parameters.gpu_sigma_r,

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

@@ -1287,6 +1287,41 @@ public class ImageDtt extends ImageDttCPU {
 			final int                 sensor_mask_inter, // The bitmask - which sensors to correlate, -1 - all.
 			final int                 threadsMax,        // maximal number of threads to launch
 			final int                 globalDebugLevel)
+	{
+		interCorrTDMotionBlur( // erase-capable version, -1 = legacy no-erase // By Claude on 07/05/2026
+				imgdtt_params, -1, tp_tasks, fcorr_td,
+				gpu_sigma_r, gpu_sigma_b, gpu_sigma_g, gpu_sigma_m,
+				gpu_sigma_rb_corr, gpu_sigma_corr, gpu_sigma_log_corr,
+				corr_red, corr_blue, sensor_mask_inter, threadsMax, globalDebugLevel);
+	}
+
+	/**
+	 * Same as interCorrTDMotionBlur() above with control of erasing the CLT before the
+	 * FIRST (positive-scale, SET) convert: erase_clt_first >0 - fill NaN, 0 - fill 0.0,
+	 * <0 - no erase (legacy). NaN-erase makes tiles absent from the task set read as NaN
+	 * downstream (imclt renders; the NaN-aware channel consolidation
+	 * CuasTD.consolidateSensorsTD - weighted-by-existing-tiles average, all-NaN where no
+	 * sensor has the tile) instead of stale data from the previous conversion. The SECOND
+	 * (negative-scale, SUBTRACT-accumulate) convert never erases.
+	 * By Claude on 07/05/2026, from Andrey's requirement.
+	 */
+	public void interCorrTDMotionBlur(
+			final ImageDttParameters  imgdtt_params,    // Now just extra correlation parameters, later will include, most others
+			final int                 erase_clt_first,  // erase CLT before the FIRST convert: >0 NaN, 0 zeros, <0 keep // By Claude on 07/05/2026
+			final TpTask[][]          tp_tasks,
+			final float  [][][][]     fcorr_td,        // [tilesY][tilesX][pair][4*64] transform domain representation of 6 corr pairs
+			final double              gpu_sigma_r,     // 0.9, 1.1
+			final double              gpu_sigma_b,     // 0.9, 1.1
+			final double              gpu_sigma_g,     // 0.6, 0.7
+			final double              gpu_sigma_m,     //  =       0.4; // 0.7;
+			final double              gpu_sigma_rb_corr,    //  = 0.5; // apply LPF after accumulating R and B correlation before G, monochrome ? 1.0 :
+			final double              gpu_sigma_corr,       //  =    0.9;gpu_sigma_corr_m
+			final double              gpu_sigma_log_corr,   // hpf to reduce dynamic range for correlations
+			final double              corr_red, // +used
+			final double              corr_blue,// +used
+			final int                 sensor_mask_inter, // The bitmask - which sensors to correlate, -1 - all.
+			final int                 threadsMax,        // maximal number of threads to launch
+			final int                 globalDebugLevel)
 	{
 		final int numcol = isMonochrome()?1:3;
 		final double [] col_weights= new double [numcol]; // colors are RBG
@@ -1346,7 +1381,7 @@ public class ImageDtt extends ImageDttCPU {
 				false); // boolean use_aux    // while is it in class member? - just to be able to free
 		
        	// Skipping if ((fdisp_dist != null) || (fpxpy != null)) {...
-		gpuQuad.execConvertDirect(false, null, -1); // Convert primary image, no erase (each tile will be SET as scales > 0
+		gpuQuad.execConvertDirect(false, null, erase_clt_first); // Convert primary image (erase per erase_clt_first; each task tile is SET as scales > 0) // By Claude on 07/05/2026
 		
 		// set secondary tasks and perform direct conversion to TD, subtracting from the converted primary
 		gpuQuad.setTasks(                  // copy tp_tasks to the GPU memory