CLAUDE: rung B - device-visible pose current/candidate set index (no host swap)

Design 3-A4i rung B (imagej-elphel-internal 2026-07-16 handoff), DP-readiness:
- gpu_pose_* / gpu_pose_candidate_* become FIXED physical slots; the CURRENT
  set is selected by a device-authoritative int flag.
- pose_lma_reduce_rms gains a final int* current_set arg and flips it on
  accept DEVICE-SIDE (NULL = no flip, standalone tests); rejection leaves it
  unchanged so the prior current raw pose remains the fallback.
- tp_proc_exec_pose_lma_resident_step: host std::swap removed; launch
  pointers come from cur_*/cand_* accessors; the host mirror syncs with a
  4-byte D2H after each step. exec_pose_fx_jacobian uploads/reads via the
  CURRENT set. The candidate-base DtoD invariant (ref rows pose-independent)
  is now documented at the copy site.
- test_pose_lma_step: pass + assert the no-flip side (not-accepted case).
C API unchanged. Gates: direct CUDA 5120/150 PASS (products 0.000244141,
candidate 9.69e-08, same as session-40); NVRTC resident accept-carryover +
worsened-rejection PASS (flip proven behaviorally); 19-float bit-exact PASS;
fx_jacobian PASS; compute-sanitizer(12.8) 0 errors; test_pose_corr_jna
bit-exact @tol 0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/tests/test_pose_lma_step.cu

@@ -253,9 +253,17 @@ int main(int argc, char ** argv)
 	checkCudaErrors(cudaGetLastError());
 	pose_lma_reduce_step<<<1, 1>>>(numPartials, lambda, pureWeight,
 			gpuPartials.dev(), gpuVectorResident.dev(), gpuResidentResult.dev());
+	// current_set: device set index (rung B) - this case reduces the SAME partials
+	// as the current step, so candidate RMS == current RMS -> NOT accepted -> the
+	// index must stay 0 (the no-flip side of the device-side commit).
+	// By Claude on 07/16/2026.
+	GpuBuf<int> gpuCurrentSet(1);
+	const int zeroSet = 0;
+	gpuCurrentSet.upload(&zeroSet);
 	pose_lma_reduce_rms<<<1, 1>>>(numPartials, pureWeight, 0.001f,
-			gpuPartials.dev(), gpuResidentResult.dev());
+			gpuPartials.dev(), gpuResidentResult.dev(), gpuCurrentSet.dev());
 	checkCudaErrors(cudaGetLastError()); checkCudaErrors(cudaDeviceSynchronize());
+	const std::vector<int> gotCurrentSet = gpuCurrentSet.download();
 	const std::vector<float> gotJt = gpuPreparedJt.download();
 	const std::vector<float> gotYmfx = gpuPreparedYmfx.download();
 	const std::vector<float> gotResident = gpuResidentResult.download();
@@ -279,7 +287,8 @@ int main(int argc, char ** argv)
 			(maxProducts <= 0.02f) && (maxDelta <= 2.0e-5f) &&
 			(maxCandidate <= 2.0e-5f) && (maxRms <= 2.0e-6f) &&
 			(gotResident[POSE_LMA_ACCEPTED] == 0.0f) &&
-			(gotResident[POSE_LMA_STOP] == 0.0f);
+			(gotResident[POSE_LMA_STOP] == 0.0f) &&
+			(gotCurrentSet[0] == 0); // not accepted -> no device-side flip // By Claude on 07/16/2026
 	std::printf("pose_lma resident 5120/150: prepared bit mismatches=%d, max products=%g, delta=%g, candidate=%g, RMS=%g -> %s\n",
 			preparedBad, maxProducts, maxDelta, maxCandidate, maxRms,
 			residentOk ? "PASS" : "FAIL");

src/tp_lma.cu

@@ -737,7 +737,8 @@ extern "C" __global__ void pose_lma_reduce_rms(
 		float         pure_weight,
 		float         rms_diff,
 		const float * partials,
-		float       * result)
+		float       * result,
+		int         * current_set) // device set index, flipped on accept (NULL = no flip) // By Claude on 07/16/2026
 {
 	if ((blockIdx.x != 0) || (threadIdx.x != 0)) return;
 	float pure_l2 = 0.0f;
@@ -760,6 +761,12 @@ extern "C" __global__ void pose_lma_reduce_rms(
 					pose_lma_sub(1.0f, rms_diff)));
 	result[POSE_LMA_ACCEPTED] = accepted ? 1.0f : 0.0f;
 	result[POSE_LMA_STOP] = stop ? 1.0f : 0.0f;
+	// Device-authoritative accept commit (rung B): flip the current/candidate
+	// buffer-set index in device memory. The host (and later a DP parent) reads
+	// this index instead of swapping pointers. By Claude on 07/16/2026.
+	if (accepted && (current_set != NULL)) {
+		*current_set ^= 1;
+	}
 }
 
 extern "C" __global__ void lma_normal_products(

src/tp_lma.h

@@ -108,12 +108,18 @@ extern "C" __global__ void pose_lma_reduce_step(
 
 // Reduce a separately projected candidate pose to RMS only. Results append to
 // the compact resident step record without changing its candidate/normal terms.
+// On accept it also flips *current_set (0<->1) DEVICE-SIDE - the authoritative
+// current/candidate buffer-set index (rung B, design 2026-07-16): a future DP
+// parent resolves the raw pose buffers with no host round trip; today the host
+// mirrors the flag with a 4-byte D2H after the step. NULL current_set = no flip
+// (standalone kernel tests). By Claude on 07/16/2026.
 extern "C" __global__ void pose_lma_reduce_rms(
 		int           num_partials,
 		float         pure_weight,
 		float         rms_diff,
 		const float * partials,
-		float       * result);        // candidate RMS fields in resident result
+		float       * result,         // candidate RMS fields in resident result
+		int         * current_set);   // device current/candidate set index, flipped on accept
 
 extern "C" __global__ void lma_normal_products(
 		int            num_params,