CLAUDE: 3-B rung B1 - lean path switches to the resident GPU task build

leanMeasure (uniform-MB lean path, JNA backend): the per-cycle CPU front
end - transformToScenePxPyD (full 5120-tile grid), setInterTasksMotionBlur
+ sort, and the two task H2D uploads - is replaced by ONE
gpuQuad.execPoseTaskUpdate() call: pose_task_update projects the
per-sequence template on-device from the Java-tracked measure pose and
rewrites both resident task slots in place; geometry + convert then run
on ACTIVATED slots with no upload (ImageDtt.interCorrTDResident). Per
cycle the only H2D left is the 12-float pose vector. Per scene:
setupERS + camera-block upload (rung C2 registers, shared with
prepare_resident) + skeleton slot re-upload (~120 KB, guards against
other pipeline stages using the task slots between scenes). The
per-scene uniform-MB 6-float descriptor replicates the exact
setInterTasksMotionBlur double crank math; margin/projection failures
become task=0 holes (missing peak -> conditioning abstains, D3 design).
Fallbacks: JCuda backend, pose_mb_uniform off, NaN/non-uniform MB
(degenerate uniformMotionBlur fallback) -> unchanged legacy CPU build.

- GpuQuad/GpuQuadJna/TpJna: execPoseTaskUpdate + activateTaskSlot
  (base returns false; JNA marshals rung-C2 nullable groups).
- IntersceneLmaFloat.buildTasks: serial float task-build oracle (the
  Java float clone of the kernel, worldFromPixel -> pixelFromWorld +
  descriptor/margin/hole packing).
- One-shot oracle at pose_lma_debug>=1: GPU stream vs the legacy Java
  DOUBLE build (txy-matched field compare, D3 gate <=1e-5 px print) and
  vs the float-serial clone; once-per-program path-active note.
- pose_corr export stays the explicit diagnostic exception: armed runs
  read the GPU-built pre-offset streams back and capture task words RAW
  (new iterTasksPre or511 overload - a |511-ed hole would diverge in a
  tol-0 replay).

mvn -DskipTests clean package: BUILD SUCCESS. Real-scene gate (D3,
ratified): poses < 0.003 px drift, aggregate RMS unchanged at print
precision, same QC count, 497/497 - Andrey's run.

CUDA side: tile_processor_gpu lwir16_2 @ ef36688 (kernel + JNA API +
two-tier tests, all regressions PASS incl. pose_corr @tol 0).
Design: internal handoffs/2026-07-16_3b_measure_chain_residency_design.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

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

@@ -28,6 +28,7 @@ import java.io.FileWriter;
 import java.io.IOException;
 import java.util.ArrayList;
 import java.util.Arrays;
+import java.util.HashMap; // B1 task-oracle txy matching // By Claude on 07/16/2026
 import java.util.concurrent.locks.LockSupport;
 
 import com.elphel.imagej.cameras.CLTParameters;
@@ -594,15 +595,30 @@ public class CuasPoseRT {
 		long measureAccounted = 0L;
 		final GpuQuad gpuQuad = center_CLT.getGPUQuad();
 		final int num_pix = pXpYD_center.length;
-		// 1. scene grid at the current pose estimate (average-camera coordinates)
+		final double disparity_corr = clt_parameters.imp.disparity_corr + center_CLT.getDispInfinityRef(); // hoisted for B1 // By Claude on 07/16/2026
 		long profileStart = (rtProfile != null) ? rtProfile.start() : 0L;
-		final double [][] scene_pXpYD = OpticalFlow.transformToScenePxPyD(
+		// 3-B rung B1 (design 2026-07-16): resident GPU task build - pose_task_update
+		// projects the per-sequence template on-device from the current pose and
+		// rewrites both native task slots in place; the full-grid transform and the
+		// CPU task build below become the fallback (JCuda backend, non-uniform MB,
+		// degenerate uniform fallback). By Claude on 07/16/2026.
+		TpTask [][] b1_tasks = null;
+		if ((mb_vectors != null) && clt_parameters.curt.pose_mb_uniform &&
+				gpuQuad.supportsResidentTaskSlots()) {
+			b1_tasks = gpuTaskBuild(clt_parameters, image_dtt, center_CLT, scene,
+					center_disparity, pXpYD_center, selection, scene_xyz, scene_atr, margin,
+					mb_vectors, mb_tau, mb_max_gain, disparity_corr, gpuQuad, debugLevel);
+		}
+		final boolean b1_built = (b1_tasks != null);
+		// 1. scene grid at the current pose estimate (average-camera coordinates) -
+		//    LEGACY CPU front end only (B1 projects on-device) // By Claude on 07/16/2026
+		final double [][] scene_pXpYD = b1_built ? null : OpticalFlow.transformToScenePxPyD(
 				null, center_disparity, scene_xyz, scene_atr, scene, center_CLT);
-		if (rtProfile != null) measureAccounted += rtProfile.addElapsed(RtPoseProfile.GRID_TRANSFORM, profileStart);
+		if (!b1_built && (rtProfile != null)) measureAccounted += rtProfile.addElapsed(RtPoseProfile.GRID_TRANSFORM, profileStart);
 		// B0 grid-transform attribution (one-shot): full-grid vs selection-only,
 		// multi-thread vs single-thread. churn ~= multi - single/threads;
 		// full-grid waste ~= full - selected-only. By Claude on 07/16/2026.
-		if ((clt_parameters.curt.pose_lma_debug >= 1) && !b0_grid_reported) {
+		if (!b1_built && (clt_parameters.curt.pose_lma_debug >= 1) && !b0_grid_reported) { // legacy-path diagnostic // By Claude on 07/16/2026
 			b0_grid_reported = true;
 			int n_def = 0;
 			for (double [] p : scene_pXpYD) if (p != null) n_def++;
@@ -626,11 +642,37 @@ public class CuasPoseRT {
 					n_def, n_sel, (b0_t1-b0_t0)*1e-6, (b0_t2-b0_t1)*1e-6,
 					(b0_t3-b0_t2)*1e-6, (b0_t4-b0_t3)*1e-6, ImageDtt.THREADS_MAX));
 		}
-		final double disparity_corr = clt_parameters.imp.disparity_corr + center_CLT.getDispInfinityRef();
 		final TpTask [][] cons_tasks; // the convert task sets, hoisted for the consolidation OOB filter // By Claude on 07/12/2026
 		final boolean resident_tasks; // post-offset slots stay native on the JNA MB path // By Codex on 07/15/2026
 		profileStart = (rtProfile != null) ? rtProfile.start() : 0L;
-		if (mb_vectors != null) {
+		if (b1_built) {
+			// 3-B rung B1: tasks are already GPU-built in the resident slots. The
+			// Java-side skeleton only carries count/txy/task_size for consolidation
+			// and the armed pose_corr export. By Claude on 07/16/2026.
+			cons_tasks = b1_tasks;
+			// armed pose_corr capture = the explicit diagnostic exception to
+			// residency: read the GPU-built PRE-OFFSET streams back and capture the
+			// task words RAW (no |511 re-OR - holes must stay 0 for a tol-0 replay).
+			if (PoseCorrExport.isArmed()) {
+				gpuQuad.updateTasks(cons_tasks[0], false, 0);
+				gpuQuad.updateTasks(cons_tasks[1], false, 1);
+				PoseCorrExport.iterTasksPre(clt_parameters, image_dtt, gpuQuad, scene, cons_tasks, false);
+			}
+			if (rtProfile != null) measureAccounted += rtProfile.addElapsed(RtPoseProfile.TASK_BUILD_SORT, profileStart);
+			profileStart = (rtProfile != null) ? rtProfile.start() : 0L;
+			image_dtt.interCorrTDResident(
+					1,                          // erase_clt_first: NaN outside the task set
+					clt_parameters.gpu_sigma_r,
+					clt_parameters.gpu_sigma_b,
+					clt_parameters.gpu_sigma_g,
+					clt_parameters.gpu_sigma_m,
+					scene.isMonochrome() ? 1.0 : clt_parameters.gpu_sigma_rb_corr,
+					clt_parameters.getGpuCorrSigma(scene.isMonochrome()),
+					clt_parameters.getGpuCorrLoGSigma(scene.isMonochrome()),
+					debugLevel);
+			if (rtProfile != null) measureAccounted += rtProfile.addElapsed(RtPoseProfile.CONVERT_GEOMETRY, profileStart);
+			resident_tasks = true;
+		} else if (mb_vectors != null) {
 			// Motion-blur compensation (lean v2, the certified CuasRender chain): two task
 			// sets, double convert (positive SET + negative shifted SUBTRACT), erase-first
 			// so undefined tiles reach downstream (consolidation, renders) as NaN.
@@ -932,6 +974,317 @@ public class CuasPoseRT {
 	private static boolean b0_grid_reported = false;
 	private static boolean b0_task_reported = false;
 	private static boolean b0_mb_reported = false;
+	// 3-B rung B1 (design 2026-07-16): resident GPU task build. By Claude on 07/16/2026.
+	private static boolean task_build_path_reported = false; // once-per-program "path active" note
+	private static boolean b1_task_oracle_reported = false;  // one-shot GPU vs Java-double vs float-serial compare
+
+	/**
+	 * Per-sequence state of the resident GPU task build (rung B1): the sorted task
+	 * template (packed txy + reference centers per selected tile), the Java-side
+	 * skeleton (carries count/txy/task_size for consolidation and the armed
+	 * pose_corr export - the live task content is GPU-resident), and the
+	 * configuration-register bookkeeping (which scene's camera block + skeleton
+	 * are resident). Keyed by the per-sequence array/object identities.
+	 * By Claude on 07/16/2026.
+	 */
+	private static final class GpuTaskTemplate {
+		final boolean []  selectionKey;
+		final double [][] centersKey;
+		final GpuQuad     gpuKey;
+		final int         count;
+		final int []      map;      // packed txy: tx | (ty<<16), (tileY,tileX)-ascending
+		final float []    centers;  // reference px,py,disparity per task
+		final TpTask [][] skeleton; // [2][count]
+		IntersceneLmaFloat.Camera refCamera = null; // captured once per sequence (after setupERS)
+		String uploadedScene = null; // scene whose camera block + skeleton slots are current
+		GpuTaskTemplate(boolean [] selectionKey, double [][] centersKey, GpuQuad gpuKey,
+				int count, int [] map, float [] centers, TpTask [][] skeleton) {
+			this.selectionKey = selectionKey;
+			this.centersKey = centersKey;
+			this.gpuKey = gpuKey;
+			this.count = count;
+			this.map = map;
+			this.centers = centers;
+			this.skeleton = skeleton;
+		}
+	}
+	private static GpuTaskTemplate gpu_task_template = null;
+
+	/**
+	 * 3-B rung B1 (design 2026-07-16): build the correlation tasks ON-DEVICE.
+	 * Verifies the MB vectors are uniform (the pose_mb_uniform broadcast), builds/
+	 * caches the per-sequence template + skeleton, refreshes the per-scene
+	 * configuration registers (scene camera block after setupERS + skeleton
+	 * re-upload, guarding against other pipeline stages having used the task
+	 * slots between scenes), computes the per-scene 6-float uniform-MB descriptor
+	 * with the EXACT setInterTasksMotionBlur double math, and launches
+	 * pose_task_update to rewrite both resident task slots from the current pose.
+	 * Zero per-cycle H2D beyond the 12-float pose vector.
+	 *
+	 * @return the Java-side skeleton TpTask[2][count] to use as cons_tasks, or
+	 *         null to fall back to the legacy CPU build (non-uniform/NaN MB).
+	 * By Claude on 07/16/2026.
+	 */
+	private static TpTask [][] gpuTaskBuild(
+			final CLTParameters clt_parameters,
+			final ImageDtt      image_dtt,
+			final QuadCLT       center_CLT,
+			final QuadCLT       scene,
+			final double []     center_disparity,
+			final double [][]   pXpYD_center,
+			final boolean []    selection,
+			final double []     scene_xyz,
+			final double []     scene_atr,
+			final int           margin,
+			final double [][]   mb_vectors,
+			final double        mb_tau,
+			final double        mb_max_gain,
+			final double        disparity_corr,
+			final GpuQuad       gpuQuad,
+			final int           debugLevel) {
+		// 1. uniform-MB verification: the descriptor is per-scene ONLY if every
+		//    selected tile carries the same (broadcast) vector; the degenerate
+		//    per-tile fallback of uniformMotionBlur() goes back to the CPU build.
+		double mb0 = Double.NaN, mb1 = Double.NaN;
+		for (int nTile = 0; nTile < selection.length; nTile++) {
+			if (!selection[nTile] || (pXpYD_center[nTile] == null)) continue;
+			final double vx = mb_vectors[0][nTile];
+			final double vy = mb_vectors[1][nTile];
+			if (Double.isNaN(vx) || Double.isNaN(vy)) return null; // NaN would drop tiles per-tile
+			if (Double.isNaN(mb0)) { mb0 = vx; mb1 = vy; }
+			else if ((vx != mb0) || (vy != mb1)) return null;      // not uniform
+		}
+		if (Double.isNaN(mb0)) return null; // no selected tiles with centers
+		// 2. per-sequence template + skeleton (cached by array/object identity)
+		GpuTaskTemplate tpl = gpu_task_template;
+		if ((tpl == null) || (tpl.selectionKey != selection) ||
+				(tpl.centersKey != pXpYD_center) || (tpl.gpuKey != gpuQuad)) {
+			final int sensor_w = scene.getGeometryCorrection().getSensorWH()[0];
+			final int tilesX = sensor_w / GPUTileProcessor.DTT_SIZE;
+			final int task_code = (1 << GPUTileProcessor.TASK_CORR_EN) | (1 << GPUTileProcessor.TASK_INTER_EN);
+			int count = 0;
+			for (int nTile = 0; nTile < selection.length; nTile++) {
+				if (selection[nTile] && (pXpYD_center[nTile] != null) &&
+						!Double.isNaN(pXpYD_center[nTile][2])) count++;
+			}
+			if (count == 0) return null;
+			final int []   map =     new int [count];
+			final float [] centers = new float [3 * count];
+			final TpTask [][] skeleton = new TpTask [2][count];
+			int i = 0;
+			for (int nTile = 0; nTile < selection.length; nTile++) { // nTile-ascending == (tileY,tileX) sort (3-A3)
+				if (!selection[nTile] || (pXpYD_center[nTile] == null) ||
+						Double.isNaN(pXpYD_center[nTile][2])) continue;
+				final int tileX = nTile % tilesX;
+				final int tileY = nTile / tilesX;
+				map[i] = tileX | (tileY << 16);
+				for (int k = 0; k < 3; k++) centers[3 * i + k] = (float) pXpYD_center[nTile][k];
+				for (int set = 0; set < 2; set++) {
+					skeleton[set][i] = new TpTask(scene.getNumSensors(), tileX, tileY);
+					skeleton[set][i].task = task_code;
+				}
+				i++;
+			}
+			center_CLT.getErsCorrection().setupERS(); // reference line tables (transformToScenePxPyD parity)
+			tpl = new GpuTaskTemplate(selection, pXpYD_center, gpuQuad,
+					count, map, centers, skeleton);
+			tpl.refCamera = IntersceneLmaFloat.Camera.capture(center_CLT.getErsCorrection());
+			gpu_task_template = tpl;
+		}
+		// 3. per-scene configuration registers: camera block (after setupERS) +
+		//    skeleton slots (re-uploaded per scene: other pipeline stages may have
+		//    replaced the task slots between scenes; ~120 KB, once per scene)
+		final boolean new_scene = !scene.getImageName().equals(tpl.uploadedScene);
+		final boolean new_template = (tpl.uploadedScene == null);
+		IntersceneLmaFloat.Camera scene_camera = null;
+		if (new_scene) {
+			scene.getErsCorrection().setupERS(); // per-scene ERS line tables (transformToScenePxPyD parity)
+			scene_camera = IntersceneLmaFloat.Camera.capture(scene.getErsCorrection());
+			gpuQuad.setTasks(tpl.skeleton[0], false, false, 0); // sizes slots + |511 task words
+			gpuQuad.setTasks(tpl.skeleton[1], false, false, 1);
+		}
+		// 4. per-scene uniform-MB descriptor - the EXACT setInterTasksMotionBlur
+		//    double math (accumulate=false, global_scale=1.0 as the leanMeasure
+		//    convenience overload passes them), cast to float at the end like
+		//    setCenterXY/setScale would
+		final int task_code_or = ((1 << GPUTileProcessor.TASK_CORR_EN) | (1 << GPUTileProcessor.TASK_INTER_EN)) | 511;
+		final double min_sub = 1e-12;
+		final double min_len = 0.001; // pix
+		double dxm, dym, dxs, dys, gain, gain_sub;
+		int task_code_sub;
+		double dx = mb0;
+		double dy = mb1;
+		double mb_len = Math.sqrt(dx * dx + dy * dy);
+		if (mb_len < min_len) { // no-MB: sub set disabled (task 0 -> |511 = 511)
+			dxm = 0.0; dym = 0.0; dxs = 0.0; dys = 0.0;
+			gain = 1.0;
+			gain_sub = -min_sub;
+			task_code_sub = 511;
+		} else {
+			dx /= mb_len; // unit vector
+			dy /= mb_len;
+			mb_len *= mb_tau; // now in pixels
+			final double mb_len_scale = -Math.log(1.0 - 1.0 / mb_max_gain);
+			double mb_offs = 1.0;
+			final double min_offs = mb_len_scale * mb_len;
+			if (mb_offs < min_offs) {
+				mb_offs = min_offs;
+			}
+			dx *= mb_offs;
+			dy *= mb_offs;
+			final double exp_offs = Math.exp(-mb_offs / mb_len);
+			final double rel_cent = mb_len * (1.0 - (mb_offs / mb_len + 1.0) * exp_offs) / (mb_offs * (1 - exp_offs));
+			dxm = -rel_cent * dx;                    // centerXY(main) = projected - rel_cent*d
+			dym = -rel_cent * dy;
+			dxs = dxm + (1 - rel_cent) * dx;         // centerXY(sub) = main + (1-rel_cent)*d (Java order preserved)
+			dys = dym + (1 - rel_cent) * dy;
+			gain = 1.0 / (1.0 - exp_offs);
+			gain_sub = -gain * exp_offs;
+			if (gain_sub > -min_sub) {
+				gain_sub = -min_sub;
+			}
+			task_code_sub = task_code_or;
+		}
+		// 5. margin gate bounds - GPU port-coords mode Java bounds (no WOI tops)
+		final int [] wh = scene.getGeometryCorrection().getSensorWH();
+		final float min_px = margin;
+		final float max_px = wh[0] - 1 - margin;
+		final float min_py = margin;
+		final float max_py = wh[1] - 1 - margin;
+		// 6. launch: template on rebuild, camera block per scene, pose every call
+		//    (48 B - the Java-tracked measure pose; B3/DP chain will drop it)
+		final float [] ref_xyz =   IntersceneLmaFloat.toFloat3(ZERO3);
+		final float [] ref_atr =   IntersceneLmaFloat.toFloat3(ZERO3);
+		final float [] scene_xyzf = IntersceneLmaFloat.toFloat3(scene_xyz);
+		final float [] scene_atrf = IntersceneLmaFloat.toFloat3(scene_atr);
+		final boolean built = gpuQuad.execPoseTaskUpdate(
+				new_template ? tpl.refCamera : null,
+				scene_camera,                     // null = keep resident (same scene)
+				ref_xyz, ref_atr, scene_xyzf, scene_atrf,
+				new_template ? tpl.map : null,
+				new_template ? tpl.centers : null,
+				tpl.count,
+				(float) disparity_corr,
+				min_px, max_px, min_py, max_py,
+				task_code_or, task_code_sub,
+				(float) dxm, (float) dym, (float) dxs, (float) dys,
+				(float) gain, (float) gain_sub);
+		if (!built) return null; // base/JCuda backend - legacy CPU build
+		if (new_scene) tpl.uploadedScene = scene.getImageName();
+		if (!task_build_path_reported) {
+			task_build_path_reported = true;
+			System.out.println("CuasPoseRT: resident GPU task build active (pose_task_update: "+
+					tpl.count+" template tasks, per-cycle H2D = 12-float pose only; "+
+					"grid transform + CPU task build + task upload eliminated on the lean path)");
+		}
+		// 7. one-shot oracle (pose_lma_debug>=1): build the tasks BOTH legacy ways
+		//    and compare - GPU vs the Java double build (the D3-ratified <=1e-5 px
+		//    field compare) and GPU vs the serial float clone (buildTasks)
+		if ((clt_parameters.curt.pose_lma_debug >= 1) && !b1_task_oracle_reported) {
+			b1_task_oracle_reported = true;
+			b1TaskOracle(clt_parameters, center_CLT, scene, center_disparity, pXpYD_center,
+					selection, scene_xyz, scene_atr, margin, mb_vectors, mb_tau, mb_max_gain,
+					disparity_corr, gpuQuad, tpl,
+					min_px, max_px, min_py, max_py, task_code_or, task_code_sub,
+					(float) dxm, (float) dym, (float) dxs, (float) dys,
+					(float) gain, (float) gain_sub);
+		}
+		return tpl.skeleton;
+	}
+
+	/**
+	 * One-shot B1 task-build oracle (pose_lma_debug>=1, first GPU-built cycle):
+	 * reads the GPU-built pre-offset streams back and compares them against
+	 * (a) the legacy Java DOUBLE build (transformToScenePxPyD +
+	 * setInterTasksMotionBlur, txy-matched - the D3-ratified <=1e-5 px field
+	 * compare) and (b) the serial FLOAT clone (IntersceneLmaFloat.buildTasks,
+	 * differences are float-scheduling only). Prints one line.
+	 * By Claude on 07/16/2026.
+	 */
+	private static void b1TaskOracle(
+			final CLTParameters clt_parameters,
+			final QuadCLT       center_CLT,
+			final QuadCLT       scene,
+			final double []     center_disparity,
+			final double [][]   pXpYD_center,
+			final boolean []    selection,
+			final double []     scene_xyz,
+			final double []     scene_atr,
+			final int           margin,
+			final double [][]   mb_vectors,
+			final double        mb_tau,
+			final double        mb_max_gain,
+			final double        disparity_corr,
+			final GpuQuad       gpuQuad,
+			final GpuTaskTemplate tpl,
+			final float min_px, final float max_px, final float min_py, final float max_py,
+			final int task_code_main, final int task_code_sub,
+			final float dxm, final float dym, final float dxs, final float dys,
+			final float scale_main, final float scale_sub) {
+		// GPU-built pre-offset streams (diagnostic readback)
+		final TpTask [] gpu0 = new TpTask [tpl.count];
+		final TpTask [] gpu1 = new TpTask [tpl.count];
+		gpuQuad.updateTasks(gpu0, false, 0);
+		gpuQuad.updateTasks(gpu1, false, 1);
+		// legacy Java DOUBLE build at the same pose
+		final double [][] scene_pXpYD = OpticalFlow.transformToScenePxPyD(
+				null, center_disparity, scene_xyz, scene_atr, scene, center_CLT);
+		final TpTask [][] cpu = GpuQuad.setInterTasksMotionBlur(
+				scene.getNumSensors(), scene.getGeometryCorrection().getSensorWH()[0],
+				false, scene_pXpYD, selection, mb_tau, mb_max_gain, mb_vectors,
+				scene.getGeometryCorrection(), disparity_corr, margin, null,
+				ImageDtt.THREADS_MAX);
+		double max_d_center = 0.0, max_d_disp = 0.0;
+		int matched = 0, gpu_only = 0, cpu_only = 0;
+		if ((cpu != null) && (cpu[0] != null)) {
+			final HashMap<Integer,Integer> cpu_index = new HashMap<>();
+			for (int j = 0; j < cpu[0].length; j++) {
+				if (cpu[0][j] != null) cpu_index.put(cpu[0][j].getTileX() | (cpu[0][j].getTileY() << 16), j);
+			}
+			for (int i = 0; i < tpl.count; i++) {
+				final boolean gpu_present = (gpu0[i].task != 0);
+				final Integer j = cpu_index.get(tpl.map[i]);
+				if (gpu_present && (j != null)) {
+					matched++;
+					max_d_center = Math.max(max_d_center, Math.abs(gpu0[i].getCenterXY()[0] - cpu[0][j].getCenterXY()[0]));
+					max_d_center = Math.max(max_d_center, Math.abs(gpu0[i].getCenterXY()[1] - cpu[0][j].getCenterXY()[1]));
+					max_d_center = Math.max(max_d_center, Math.abs(gpu1[i].getCenterXY()[0] - cpu[1][j].getCenterXY()[0]));
+					max_d_center = Math.max(max_d_center, Math.abs(gpu1[i].getCenterXY()[1] - cpu[1][j].getCenterXY()[1]));
+					max_d_disp = Math.max(max_d_disp, Math.abs(gpu0[i].target_disparity - cpu[0][j].target_disparity));
+				} else if (gpu_present) gpu_only++;
+				else if (j != null) cpu_only++;
+			}
+		}
+		// serial FLOAT clone of the kernel
+		final float [][] oracle = IntersceneLmaFloat.buildTasks(
+				tpl.refCamera, IntersceneLmaFloat.Camera.capture(scene.getErsCorrection()),
+				IntersceneLmaFloat.toFloat3(ZERO3), IntersceneLmaFloat.toFloat3(ZERO3),
+				IntersceneLmaFloat.toFloat3(scene_xyz), IntersceneLmaFloat.toFloat3(scene_atr),
+				tpl.map, tpl.centers, TpTask.getSize(scene.getNumSensors()),
+				(float) disparity_corr, min_px, max_px, min_py, max_py,
+				task_code_main, task_code_sub, dxm, dym, dxs, dys, scale_main, scale_sub);
+		final int task_size = TpTask.getSize(scene.getNumSensors());
+		double max_d_float = 0.0;
+		int hole_mismatch = 0;
+		for (int i = 0; i < tpl.count; i++) {
+			final int base = i * task_size;
+			final boolean gpu_present = (gpu0[i].task != 0);
+			final boolean oracle_present = (Float.floatToRawIntBits(oracle[0][base]) != 0);
+			if (gpu_present != oracle_present) { hole_mismatch++; continue; }
+			if (!gpu_present) continue;
+			max_d_float = Math.max(max_d_float, Math.abs(gpu0[i].getCenterXY()[0] - oracle[0][base + 3]));
+			max_d_float = Math.max(max_d_float, Math.abs(gpu0[i].getCenterXY()[1] - oracle[0][base + 4]));
+			max_d_float = Math.max(max_d_float, Math.abs(gpu1[i].getCenterXY()[0] - oracle[1][base + 3]));
+			max_d_float = Math.max(max_d_float, Math.abs(gpu1[i].getCenterXY()[1] - oracle[1][base + 4]));
+			max_d_float = Math.max(max_d_float, Math.abs(gpu0[i].target_disparity - oracle[0][base + 2]));
+		}
+		System.out.println(String.format(
+				"CuasPoseRT B1 task oracle (one-shot): %d template tasks; GPU vs Java-double:"+
+				" matched=%d max|dCenterXY|=%.3g px max|dDisp|=%.3g (gate <=1e-5), gpu-only=%d cpu-only=%d;"+
+				" GPU vs float-serial: max|d|=%.3g px, set-mismatch=%d",
+				tpl.count, matched, max_d_center, max_d_disp, gpu_only, cpu_only,
+				max_d_float, hole_mismatch));
+	}
 
 	private static int GPUTileProcessorDttSize() {
 		return com.elphel.imagej.gpu.GPUTileProcessor.DTT_SIZE;

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

@@ -2984,6 +2984,51 @@ public class GpuQuad{ // quad camera description
 		return null;
 	}
 
+	/**
+	 * 3-B rung B1 (design 2026-07-16): resident task build - pose_task_update runs
+	 * the C1-validated projection chain over the per-sequence task template and
+	 * rewrites the 6-float task headers of BOTH resident slots in place (slot 0 =
+	 * MB main, slot 1 = MB partner; margin/projection failures become task=0
+	 * holes). Nullable groups follow the rung-C2 configuration-register contract.
+	 * Task slots must hold the per-sequence skeleton (setTasks to slots 0 and 1)
+	 * with exactly num_tasks entries. Base/JCuda returns false so the Java task
+	 * build remains the fallback. By Claude on 07/16/2026.
+	 */
+	public boolean execPoseTaskUpdate(
+			IntersceneLmaFloat.Camera reference, // null = keep resident
+			IntersceneLmaFloat.Camera scene,     // null = keep resident
+			float [] reference_xyz,              // pose group: all four null = device current set
+			float [] reference_atr,
+			float [] scene_xyz,
+			float [] scene_atr,
+			int []   task_map,                   // null = keep resident template
+			float [] task_centers,               // null = keep resident template
+			int      num_tasks,
+			float    disparity_corr,
+			float    min_px,
+			float    max_px,
+			float    min_py,
+			float    max_py,
+			int      task_code_main,             // pre-encoded (|511) task words
+			int      task_code_sub,
+			float    dx_main,                    // uniform-MB descriptor (offsets from the projected point)
+			float    dy_main,
+			float    dx_sub,
+			float    dy_sub,
+			float    scale_main,                 // pre-encoded scale fields (TpTask.setScale semantics)
+			float    scale_sub) {
+		return false;
+	}
+
+	/**
+	 * 3-B rung B1: make a resident task slot the live kernel stream WITHOUT an
+	 * H2D upload (after execPoseTaskUpdate rewrote it in place). Base/JCuda
+	 * returns false (no resident slots). By Claude on 07/16/2026.
+	 */
+	public boolean activateTaskSlot(int slot) {
+		return false;
+	}
+
 	public void execRBGA(
 			double [] color_weights,
 			boolean   is_lwir,

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

@@ -634,6 +634,83 @@ public class GpuQuadJna extends GpuQuad {
                 outY, outWeights, outEigen);
     }
 
+    // ---- 3-B rung B1: resident task build (design 2026-07-16) ---- // By Claude on 07/16/2026
+    @Override public boolean execPoseTaskUpdate(
+            final IntersceneLmaFloat.Camera reference, // null = keep resident
+            final IntersceneLmaFloat.Camera scene,     // null = keep resident
+            final float[] reference_xyz,               // all four null = keep device current set
+            final float[] reference_atr,
+            final float[] scene_xyz,
+            final float[] scene_atr,
+            final int[]   task_map,                    // null = keep resident template
+            final float[] task_centers,                // null = keep resident template
+            final int     num_tasks,
+            final float   disparity_corr,
+            final float   min_px,
+            final float   max_px,
+            final float   min_py,
+            final float   max_py,
+            final int     task_code_main,
+            final int     task_code_sub,
+            final float   dx_main,
+            final float   dy_main,
+            final float   dx_sub,
+            final float   dy_sub,
+            final float   scale_main,
+            final float   scale_sub) {
+        if ((num_tasks <= 0) ||
+                ((task_map != null) != (task_centers != null)) ||
+                ((task_map != null) && ((task_map.length != num_tasks) ||
+                        (task_centers.length != 3 * num_tasks)))) {
+            throw new IllegalArgumentException("GpuQuadJna.execPoseTaskUpdate: inconsistent inputs");
+        }
+        final boolean have_pose = (scene_xyz != null) || (scene_atr != null) ||
+                (reference_xyz != null) || (reference_atr != null);
+        float[] poseVectors = null;
+        if (have_pose) {
+            poseVectors = new float[12];
+            copyPoseVector3(reference_xyz, poseVectors, 0);
+            copyPoseVector3(reference_atr, poseVectors, 3);
+            copyPoseVector3(scene_xyz, poseVectors, 6);
+            copyPoseVector3(scene_atr, poseVectors, 9);
+        }
+        final int rc = lib.tp_proc_exec_pose_task_update(
+                proc,
+                (reference != null) ? poseCameraMeta(reference) : null,
+                (reference != null) ? reference.radial : null,
+                (reference != null) ? reference.rByRDist : null,
+                (reference != null) ? reference.rByRDist.length : 0,
+                (reference != null) ? reference.ers : null,
+                (reference != null) ? reference.ers.length : 0,
+                (scene != null) ? poseCameraMeta(scene) : null,
+                (scene != null) ? scene.radial : null,
+                (scene != null) ? scene.rByRDist : null,
+                (scene != null) ? scene.rByRDist.length : 0,
+                (scene != null) ? scene.ers : null,
+                (scene != null) ? scene.ers.length : 0,
+                poseVectors, task_map, task_centers,
+                num_tasks, getNumSensors(),
+                disparity_corr,
+                min_px, max_px, min_py, max_py,
+                task_code_main, task_code_sub,
+                dx_main, dy_main, dx_sub, dy_sub,
+                scale_main, scale_sub);
+        if (rc != 0) {
+            throw new IllegalStateException("GpuQuadJna.execPoseTaskUpdate rc="+rc+": "+
+                    lib.tp_last_error());
+        }
+        return true;
+    }
+
+    @Override public boolean activateTaskSlot(final int slot) { // By Claude on 07/16/2026
+        final int rc = lib.tp_proc_activate_tasks_slot(proc, slot);
+        if (rc != 0) {
+            throw new IllegalStateException("GpuQuadJna.activateTaskSlot("+slot+") rc="+rc+": "+
+                    lib.tp_last_error());
+        }
+        return true;
+    }
+
     // ---- resident float pose preparation/reduction continuation (roadmap rung 3-A4e) ----
     @Override public IntersceneLmaFloat.ResidentLmaStepResult execPoseLmaResidentStep(
             final float lambda,

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

@@ -89,6 +89,9 @@ public interface TpJna extends Library {
     /** Explicit persistent task slots (0/1); the selected slot becomes the live kernel stream. */
     int     tp_proc_set_tasks_slot(Pointer proc, float[] ftasks, int ntiles, int totalFloats, int slot);
     int     tp_proc_get_tasks_slot(Pointer proc, float[] out, int totalFloats, int slot);
+    /** 3-B rung B1: make a resident slot the live kernel task stream with NO upload
+     *  (after pose_task_update rewrote it in place). By Claude on 07/16/2026. */
+    int     tp_proc_activate_tasks_slot(Pointer proc, int slot);
     int     tp_proc_exec_geometry(Pointer proc, int uniformGrid);
     /** convert_direct: ref_scene (0/1 -> clt vs clt_ref), erase_clt (-1/0/1), no_kernels (0/1). */
     int     tp_proc_exec_convert_direct(Pointer proc, int refScene, int eraseClt, int noKernels);
@@ -182,6 +185,26 @@ public interface TpJna extends Library {
                                       int light, // rung C3: 1 = frozen conditioning, fresh y only // By Claude on 07/17/2026
                                       float[] prepResult, int capture,
                                       float[] outY, float[] outWeights, float[] outEigen);
+    /** 3-B rung B1 (design 2026-07-16): resident task build. pose_task_update runs the
+     *  C1-validated projection chain over the per-sequence task template and rewrites
+     *  the 6-float task headers of BOTH resident slots in place. Nullable groups follow
+     *  the rung-C2 configuration-register contract: camera blocks (shared with
+     *  prepare_resident), taskMap/taskCenters (per-sequence template), poseVectors
+     *  (null = device-authoritative current set). Task slots must hold the
+     *  per-sequence skeleton (set_tasks_slot) with exactly numTasks entries.
+     *  By Claude on 07/16/2026. */
+    int     tp_proc_exec_pose_task_update(Pointer proc,
+                                      float[] refMeta, float[] refRadial,
+                                      float[] refRbr, int refRbrLen, float[] refErs, int refErsLen,
+                                      float[] sceneMeta, float[] sceneRadial,
+                                      float[] sceneRbr, int sceneRbrLen, float[] sceneErs, int sceneErsLen,
+                                      float[] poseVectors, int[] taskMap, float[] taskCenters,
+                                      int numTasks, int numCams,
+                                      float disparityCorr,
+                                      float minPx, float maxPx, float minPy, float maxPy,
+                                      int taskCodeMain, int taskCodeSub,
+                                      float dxMain, float dyMain, float dxSub, float dySub,
+                                      float scaleMain, float scaleSub);
     /** Deterministic double normal-equation products for the lean pose LMA.
      *  jt is parameter-major [numParams][numValues]; out is row-major H followed by b.
      *  Damping and the small solve remain in Java. 0 on success. */

src/main/java/com/elphel/imagej/gpu/testdata/PoseCorrExport.java

@@ -184,6 +184,24 @@ public class PoseCorrExport {
 			GpuQuad       gpuQuad,
 			QuadCLT       scene,
 			TpTask [][]   tp_tasks) {
+		iterTasksPre(clt_parameters, image_dtt, gpuQuad, scene, tp_tasks, true);
+	}
+
+	/**
+	 * 3-B rung B1 overload: or511=false captures the task words RAW. On the
+	 * GPU-task-build path the pre streams are read back from the resident slots
+	 * (updateTasks D2H) - the |511 OR already happened when the per-sequence
+	 * skeleton was uploaded, and task=0 HOLES must stay 0 so a tol-0 replay
+	 * uploads exactly what the production GPU held (a |511-ed hole would get
+	 * converted by index_direct and diverge). By Claude on 07/16/2026.
+	 */
+	public static synchronized void iterTasksPre(
+			CLTParameters clt_parameters,
+			ImageDtt      image_dtt,
+			GpuQuad       gpuQuad,
+			QuadCLT       scene,
+			TpTask [][]   tp_tasks,
+			boolean       or511) {
 		if (!armed || done) return;
 		if (bound_scene == null) { // bind to this scene
 			if ((images_scene == null) || !images_scene.equals(scene.getImageName())) {
@@ -200,8 +218,8 @@ public class PoseCorrExport {
 		} else if (!bound_scene.equals(scene.getImageName())) {
 			return; // another scene (bound scene failed mid-fit and was reported) - keep what we have
 		}
-		tasks0_pre.add(flattenTasks(tp_tasks[0], true));
-		tasks1_pre.add((tp_tasks.length > 1) && (tp_tasks[1] != null) ? flattenTasks(tp_tasks[1], true) : null);
+		tasks0_pre.add(flattenTasks(tp_tasks[0], or511));
+		tasks1_pre.add((tp_tasks.length > 1) && (tp_tasks[1] != null) ? flattenTasks(tp_tasks[1], or511) : null);
 	}
 
 	/**

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

@@ -1450,6 +1450,63 @@ public class ImageDtt extends ImageDttCPU {
 		return;
 	}
 
+	/**
+	 * 3-B rung B1 (design 2026-07-16): the interCorrTDMotionBlur convert flow for
+	 * tasks that are ALREADY resident - pose_task_update rewrote both native task
+	 * slots on-device, so there is no per-cycle task H2D: each slot is only
+	 * ACTIVATED (no upload), then geometry + convert run exactly as in
+	 * interCorrTDMotionBlur (slot 0 = MB main, erase per erase_clt_first, SET;
+	 * slot 1 = MB partner, no erase, SUBTRACT). LPF constant setup is kept
+	 * identical to the legacy flow. Convert-only (the leanMeasure caller
+	 * correlates through the consolidation bridge). Requires a backend with
+	 * resident task slots (GpuQuad.activateTaskSlot). By Claude on 07/16/2026.
+	 */
+	public void interCorrTDResident(
+			final int                 erase_clt_first,
+			final double              gpu_sigma_r,
+			final double              gpu_sigma_b,
+			final double              gpu_sigma_g,
+			final double              gpu_sigma_m,
+			final double              gpu_sigma_rb_corr,
+			final double              gpu_sigma_corr,
+			final double              gpu_sigma_log_corr,
+			final int                 globalDebugLevel)
+	{
+		final float [][] lpf_rgb = new float[][] {
+			floatGetCltLpfFd(gpu_sigma_r),
+			floatGetCltLpfFd(gpu_sigma_b),
+			floatGetCltLpfFd(gpu_sigma_g),
+			floatGetCltLpfFd(gpu_sigma_m)
+		};
+		gpuQuad.setLpfRbg( // constants memory - same for all cameras
+				lpf_rgb,
+				globalDebugLevel > 2);
+		final float [] lpf_flat = floatGetCltLpfFd(gpu_sigma_corr);
+		gpuQuad.setLpfCorr(
+				"lpf_corr",
+				lpf_flat,
+				globalDebugLevel > 2);
+		final float [] lpf_rb_flat = floatGetCltLpfFd(gpu_sigma_rb_corr);
+		gpuQuad.setLpfCorr(
+				"lpf_rb_corr",
+				lpf_rb_flat,
+				globalDebugLevel > 2);
+		final float [] log_flat = floatGetCltHpfFd(gpu_sigma_log_corr);
+		gpuQuad.setLpfCorr(
+				"LoG_corr",
+				log_flat,
+				globalDebugLevel > 2);
+		// primary set (MB main): activate resident slot 0, offsets, convert-SET
+		if (!gpuQuad.activateTaskSlot(0)) {
+			throw new IllegalStateException("interCorrTDResident: backend has no resident task slots");
+		}
+		gpuQuad.execSetTilesOffsets(false); // NON-UNIFORM grid (centerXY from pose_task_update)
+		gpuQuad.execConvertDirect(false, null, erase_clt_first);
+		// secondary set (MB partner): activate resident slot 1, offsets, convert-SUBTRACT
+		gpuQuad.activateTaskSlot(1);
+		gpuQuad.execSetTilesOffsets(false);
+		gpuQuad.execConvertDirect(false, null, -1);
+	}
 
 
 	/**

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

@@ -793,6 +793,96 @@ public final class IntersceneLmaFloat {
 		return new Result(numTiles, fx, jt, valid, numValid);
 	}
 
+	/**
+	 * 3-B rung B1 (design 2026-07-16): serial float task-build oracle - the Java
+	 * float clone of the pose_task_update CUDA kernel. Runs the SAME float
+	 * projection chain as {@link #getFxDerivs} (worldFromPixel -> pixelFromWorld)
+	 * over the per-sequence task template and packs the 6-float task headers of
+	 * both MB sets exactly as the kernel does: centerXY = projected point + the
+	 * uniform-MB descriptor offsets, target_disparity = scene disparity +
+	 * disparity_corr, pre-encoded task/scale words, task=0 HOLES (all header
+	 * floats zero except txy) for failed projections and margin-gate rejects.
+	 * Differences vs the CUDA stream are float-scheduling only (FMA contraction);
+	 * the one-shot leanMeasure compare prints the max field deltas.
+	 * @return float[2][num_tasks*task_size] header-filled task streams
+	 *         (set 0 = MB main, set 1 = MB partner); xy/disp_dist tails zero.
+	 * By Claude on 07/16/2026.
+	 */
+	public static float [][] buildTasks(
+			final Camera reference,
+			final Camera scene,
+			final float [] referenceXyz,
+			final float [] referenceAtr,
+			final float [] sceneXyz,
+			final float [] sceneAtr,
+			final int []   taskMap,      // packed txy: tx | (ty<<16)
+			final float [] taskCenters,  // reference px,py,disparity per task
+			final int taskSize,
+			final float disparityCorr,
+			final float minPx,
+			final float maxPx,
+			final float minPy,
+			final float maxPy,
+			final int taskCodeMain,
+			final int taskCodeSub,
+			final float dxMain,
+			final float dyMain,
+			final float dxSub,
+			final float dySub,
+			final float scaleMain,
+			final float scaleSub) {
+		if ((taskMap == null) || (taskCenters == null) ||
+				(taskCenters.length != 3 * taskMap.length) || (taskSize < 6)) {
+			throw new IllegalArgumentException("buildTasks: inconsistent template");
+		}
+		checkVector3(referenceXyz, "referenceXyz");
+		checkVector3(referenceAtr, "referenceAtr");
+		checkVector3(sceneXyz, "sceneXyz");
+		checkVector3(sceneAtr, "sceneAtr");
+		final int numTasks = taskMap.length;
+		final float [] refInverse = rotationMatrices(referenceAtr, true);
+		final float [] sceneDirect = rotationMatrices(sceneAtr, false);
+		final float [][] streams = new float [2][numTasks * taskSize];
+		for (int task = 0; task < numTasks; task++) {
+			final int base = task * taskSize;
+			streams[0][base + 1] = Float.intBitsToFloat(taskMap[task]); // txy in both sets
+			streams[1][base + 1] = streams[0][base + 1];
+			final int ci = 3 * task;
+			final float px = taskCenters[ci];
+			final float py = taskCenters[ci + 1];
+			final float disparity = taskCenters[ci + 2];
+			if (!finite(px) || !finite(py) || !finite(disparity)) continue; // hole
+			final boolean infinity = disparity < INFINITY_DISPARITY;
+			final WorldDerivs referenceWorld = worldFromPixel(
+					reference, px, py, disparity, infinity, referenceXyz, refInverse);
+			if (referenceWorld == null) continue; // hole
+			final float [] scenePixel = pixelFromWorld(
+					scene, referenceWorld.world, infinity, sceneXyz, sceneDirect);
+			if (scenePixel == null) continue; // hole
+			final float cxMain = scenePixel[0] + dxMain;
+			final float cyMain = scenePixel[1] + dyMain;
+			final float cxSub =  scenePixel[0] + dxSub;
+			final float cySub =  scenePixel[1] + dySub;
+			if ((cxMain < minPx) || (cxMain > maxPx) || (cyMain < minPy) || (cyMain > maxPy) ||
+					(cxSub < minPx) || (cxSub > maxPx) || (cySub < minPy) || (cySub > maxPy)) {
+				continue; // margin-gate hole
+			}
+			final float targetDisparity = scenePixel[2] + disparityCorr;
+			if (!finite(targetDisparity)) continue; // hole
+			streams[0][base] =     Float.intBitsToFloat(taskCodeMain);
+			streams[1][base] =     Float.intBitsToFloat(taskCodeSub);
+			streams[0][base + 2] = targetDisparity;
+			streams[1][base + 2] = targetDisparity;
+			streams[0][base + 3] = cxMain;
+			streams[0][base + 4] = cyMain;
+			streams[1][base + 3] = cxSub;
+			streams[1][base + 4] = cySub;
+			streams[0][base + 5] = scaleMain;
+			streams[1][base + 5] = scaleSub;
+		}
+		return streams;
+	}
+
 	/** Compare a float result with the exact legacy double geometry call. */
 	public static Comparison compareToDouble(
 			final Result result,