Design 3-A4i rung C1: the LMA measured inputs never leave the GPU.
- pose_lma_assemble (grid over resident corr rows): peaks {dx,dy,str,e0x,
  e0y,l0,l1,valid} + packed corr indices -> per-tile y offsets, raw strength
  weights, eigen 2x2 transform (setEigenTransform math, float; identity when
  use_eigen==0) and the fx selection mask. Parameterized contract (counts,
  slot, shift, components, policy scalars) - CUAS specifics stay host policy.
- pose_lma_prepare_norm + pose_lma_prepare_finish: SERIAL one-thread tails
  (sample normalize + pull rows; y+=fx, H diagonal in Java getWJtJlambda
  order, regularization weights, full-weight normalize + pure_weight) -
  serial by design so the Java/C++ float clones gate BIT-EXACT.
- tp_proc_exec_pose_lma_prepare_resident: camera/pose/centers/policy H2D,
  assemble -> norm -> pose_fx_jacobian(current, NO readback) -> finish;
  D2H = 8-float compact result (+optional capture of y/weights/eigen for
  the one-shot oracle). Same scratch bookkeeping as the step so a following
  no-H2D step never regrows.
- tp_proc_exec_pose_lma_resident_step: all-NULL weights/y/eigen = use the
  resident prepared buffers (skip H2D); mixed NULLs rejected.
Gates: direct CUDA prepare 5120/159 BIT-EXACT vs serial float oracle
(y/weights/eigen/selection/scalars all 0 mismatches); resident + 19-float +
fx_jacobian NVRTC PASS; pose_corr bit-exact @tol 0; compute-sanitizer 0
errors on both binaries.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

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>

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

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

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

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

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

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

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

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

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

expected_gpu_peaks_it<k> (exported once the production hookup runs, imagej
f8df8dce) rows align with expected_corr_indices; the float-kernel replay is
compared bit-exactly at --tol, matched by packed index. Absent buffer =
skipped (pre-hookup cases). Current case: ALL PASS unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CLAUDE: peak kernel precision two-step - templated body, float instantiation corr2D_peak_eig_f (Andrey's methodology)

Step 1 (double vs double Java oracle) was the validated state; the body is
now a template on the accumulator type with two extern-C instantiations:
corr2D_peak_eig (double, the oracle-matched reference) and corr2D_peak_eig_f
(float, ~64x cheaper FLOPs + half the shared footprint - kept adequate for
full-frame, Andrey's ruling). tp_proc_exec_corr2d_peak grows a use_float arg;
NVRTC module = 25 kernels. All three tests run BOTH and report float-vs-double
grouped stats (dx/dy px, strength rel, eigvec abs, lambda rel).

Step 2 measured:
- synthetic: dx/dy <= 4.8e-7 px, lambda <= 9.4e-7 rel; eigvec 0.031 ONLY on
  the (near-)isotropic tiles where the direction is ill-defined (float can't
  fire the oracle's e=1e-12 branch) AND unused (setEigenTransform degenerates
  to isotropic weighting when lambda0~lambda1) - documented in tp_peak.cu.
- REAL case (4 iterations x 150 tiles): dx/dy <= 7.2e-7 px, strength
  IDENTICAL, eigvec <= 8.9e-5, lambda <= 1.2e-6 rel - 4+ orders below the
  ~0.01 px measurement noise. Float is adequate for production; double stays
  as the validation/oracle tier. Double regression unchanged (synthetic 0 /
  9.6e-17; real case ALL PASS).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

First rung-2 real case (Andrey's 07/14 export) measured the predicted tier
precisely: tasks + corr_td + corr_pd ALL BIT-EXACT @tol 0 (the in-place
FZ-norm is production-identical); peaks ~14% of elements flip by 1-2 float
ULP at their own magnitude (double-math FMA contraction vs Java, absorbed
except at float cast boundaries) - a flat absolute tol cannot cover 1 ULP
of lambda~40 (3.8e-6) while staying meaningful for dx~0.01. Criterion now:
pass if within --peak-tol ABSOLUTE (near-zero floor, default 1e-6) OR
within 4 float ULP of the operand (measured max 2 ULP; near-zero relative
outliers ~21 ULP are 3e-9 absolute = caught by the floor). Verdict on the
fresh case: ALL PASS, viol=0.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The scene-sequence .list is the single source of truth: it now carries
SET casesFile lists/cases.list (relative to rootDirectory, the
EyesisCorrectionParameters rule). run_cases.sh discovery order: -c arg ->
the scene list's SET casesFile (when -x/-l provides the list; explicit
pointer, missing target = hard error) -> convention lookup under
rootDirectory -> repo-local gitignored link. Verified: -l, -x and bare
modes all resolve to the same manifest/model.

Java companion (imagej-elphel): casesFile added to KEY_DIRS (index 12,
file-exempt from the auto-mkdirs) so the SET parses cleanly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Andrey's rulings 07/13/2026: data directories are not git-controlled; the
repo is shared across computers with different data sets, so it carries
code, generators and a heavily-commented TEMPLATE (linux-config style) -
everything data-specific lives under the data root:

- cases.list -> cases.list.tmpl (tracked; install by copying to
  <dataRoot>/lists/cases.list next to the scene lists; SET listFile may
  then be a bare filename - resolved relative to the manifest's directory).
- /cases.list gitignored: optional repo-local symlink into the data root
  so a bare ./run_cases.sh works.
- run_cases.sh: manifest lookup order -c arg -> ./cases.list link ->
  via -x/-l through the scene list's rootDirectory (<root>/lists/,
  <root>/, next to the scene list). -x corr-xml alone now finds
  everything (list, version, manifest) from the entries the Java run uses.

Verified live: installed at /home/elphel/lwir16-proc/LV/lists/cases.list;
bare run ALL PASS via the symlink; -x corr-xml discovery ALL resolved with
the symlink removed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CLAUDE: manifest-like tests top - cases.list + run_cases.sh (Andrey's design), de-hardwire test defaults

The scene-sequence .list is the human single source of truth (SET header +
body lines selected by uncommenting; commented lines document the setup).
The tests now follow the SAME convention one level up:

- cases.list: SET header (listFile, version - the version deliberately kept
  out of the scene list, mirroring CORRECTION_PARAMETERS.x3dModelVersion vs
  .sourceSequencesList) + case lines selected by uncommenting; commented
  cases stay as documentation. gen lines map synthetic cases to generators.
- run_cases.sh: resolves ${MODEL} through the scene list's SET
  rootDirectory/x3dDirectory + its UNCOMMENTED sequence lines + the version
  (the *-CENTER dir carries the DERIVED center timestamp -> matched as the
  smallest center ts >= the sequence's first-scene ts containing <version>/);
  -x <corr-xml> reads sourceSequencesList/x3dModelVersion directly = the
  same two sources that drive the Java run. No args = all uncommented cases;
  named cases run even when commented; missing synthetic data triggers the
  generator.
- test_proc_consolidate: hardwired absolute case/repo paths -> standard
  --data/--src/--devrt interface (legacy positional kept).
- test_pose_corr_jna: srcdir default de-hardwired (repo-relative 'src').

Verified live: ./run_cases.sh ALL PASS (pose_corr real-scene tol-0 resolved
through the list+version, 4 synthetic cases); named-commented case
avg_td_oob_jna PASS bit-exact via the new interface; gen-on-missing works.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CLAUDE: 3-C rung 2 - corr2D_peak_eig (frozen 15-arg getMaxXYCmEig contract) + checkpoint-2 replay + synthetic numpy gate

Andrey's ruling 07/13/2026: the GPU peak contract = the 15-arg superset -
closed-form 2x2 eigen, recentering as a runtime param, envelope de-bias as
an optional array pointer (NULL = off).

- src/tp_peak.{h,cu}: corr2D_peak_eig, one block/tile - parallel load/NaN/
  argmax(first-occurrence)/8-connected grow, thread-0 serial DOUBLE centroid/
  covariance/eigen/recentering in the Java oracle's exact summation order.
- jna/tp_jna.cpp: kernel in the NVRTC module (24 kernels) +
  tp_proc_exec_corr2d_peak / tp_proc_get_peaks / tp_proc_set_peak_debias.
- test_pose_corr + test_pose_corr_jna: checkpoint-2 replay (in-place
  corr2D_normalize -> corr_pd tol tier, peaks vs the double Java oracle at
  --peak-tol) when the case carries it; pre-rung-2 cases skip with a note.
- make_peak_case.py + test_peak: synthetic gate vs an INDEPENDENT numpy
  float64 port (8 adversarial tiles + refine/debias case) - PASS at
  max|diff| 0 / 9.6e-17.
- Regressions: test_pose_corr_jna bit-exact @tol 0 on the existing v013 case,
  avg_td / avg_td_oob / convert_direct PASS.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CLAUDE: test_pose_corr_jna - the tol-0 verdict binary (NVRTC production module); direct test = stepping companion

RUNG 1 VALIDATED on the first real v013 case (scene 1773135467_917220,
4 LMA iterations, 150 tiles/set, 300 corr tiles/iter): test_pose_corr_jna
PASS BIT-EXACT @tol 0 - tasks (calculate_tiles_offsets) and raw TD
correlation (correlate2D_inter) both max|diff|=0, every iteration.

The offline-nvcc direct-launch binary (src/tests/test_pose_corr) cannot be
bit-exact vs the NVRTC JIT (FMA/codegen divergence), measured: -G -O0 ->
tasks off by 2^-13 px; RELEASE=1 -> tasks BIT-EXACT, corr within ~1 ULP at
operand scale (max|diff| 128..256 vs amplitudes ~2e9, float spacing 256).
Role split documented in both headers (same pairing as test_avg_td_oob /
test_proc_consolidate): direct binary for cuda-gdb stepping, _jna binary
for the production-equivalence verdict. Build line in the _jna header.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Standalone per-iteration replay of the CuasPoseRT.leanMeasure GPU chain
against the PoseCorrExport case (imagej-elphel curt.kernel_test=pose_corr):
calcReverseDistortionTable + calc_rot_deriv + calculate_tiles_offsets (NEW
under test; GPU-filled tasks compared bit-exact vs the oracle post-offsets
streams, header+xy fields - the disp_dist tail is not round-tripped by the
Java flatten), erase_clt_tiles(NaN) + convert_direct for both MB sets,
the v1 consolidation chain into the cam-0 slot (inputs = the oracle
post-offsets streams, the exact bytes production flattens), then
correlate2D_inter vs the resident center TD from gpu_clt_ref cam 0 (NEW
under test). Raw-TD compare is ORDER-INDEPENDENT (corr slot placement is
atomicAdd-nondeterministic): tiles matched by packed index, tol 0 goal;
results/corr_td_it<k> saved in oracle order for the Java round trip. All
four LPF/HPF __constant__ tables overwritten from the case via extern
__constant__ + cudaMemcpyToSymbol (correlate2D_inter reads LoG_corr and
lpf_rb_corr). Built by build_tests.sh (auto-discovered), cuda-gdb steppable.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Wire the frozen v1 chain (index_consolidate -> consolidate_oob ->
clt_average_sensors_list) into the production JNA surface (roadmap 2c ->
step 4 JNA hookup, Andrey's ruling: JNA-only, no JCuda launch code):
- tp_jna.cpp: tp_consolidate.h/.cu added to the NVRTC module (23 kernels);
  new export tp_proc_exec_consolidate = host orchestration of the chain on
  the RESIDENT per-cam CLT (contiguous DtoD gather, no D2H/H2D of TD data;
  only the flattened task streams cross the boundary); average lands in the
  cam-0 slot (what the single conj-multiply correlates); lazy scratch,
  freed in tp_proc_destroy; stats {pairs, surviving tiles, misaligned}.
- tp_consolidate.h/.cu: JCUDA include guards (NVRTC concatenation has no
  include paths) + NAN/INFINITY fallbacks (NVRTC has no math.h macros).
- jna/test_proc_consolidate.cu: standalone validation of the export vs the
  real-scene avg_td_oob oracle case - PASS bit-exact (max|diff|=0,
  5120 pairs -> 5056 tiles == oracle, 0 misaligned).
Regressions: tests_bin/test_avg_td, test_avg_td_oob PASS at tol 0 after
the guard changes; module smoke = 23/23 kernels resolved.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Three kernels (tp_consolidate.cu, oracle = imagej-elphel CuasTD @eab1660e):
- index_consolidate: index_direct successor aware of the motion-blur task
  pair (gpu_ftasks1 may be NULL = no-MB, mirrors Java TpTask[2][] with [1]
  null); keeps tiles active in BOTH sets, txy-alignment verified
  (misaligned counted, skipped); output = compacted even/odd PAIRS (no-MB:
  odd = copy) so pass 2 is uniform. Activity tested on the task word as
  int (v0 index_direct float-compare -0.0 quirk noted).
- consolidate_oob: frozen 2c spec - per sensor depth = min over the MB
  pair of tile-center distance to nearest frame edge; <hard(8) always
  skipped, any >=soft(12) only those, else hard survivors last resort,
  nobody -> tile dropped; NaN xy excluded (explicit isnan - fminf would
  silently ignore NaN); mask folded into task bits TASK_AVG_SENS_SHIFT=11
  (bits 11+ virgin per the TpTask bit-field survey).
- clt_average_sensors_list: v0 core driven by the compacted list, gated
  by the mask; host prefills NaN/0 = oracle poison semantics.
v0 clt_average_sensors refactored onto the shared consolidate_chunk core
(all-ones mask), verified bit-exact.

test_avg_td_oob: loads AvgTdExport.exportOob cases (raw ftasks0/1
streams); runs the chain; compares masks exactly + td_avg/counts at tol 0.
make_avg_td_oob_case.py: synthetic 4-sensor case = the SAME scenario the
Java oracle self-test passes (every OOB branch); independent float32
oracle-op-order expectations.
VERIFIED: synth avg_td (v0 refactor) max|diff|=0; synth avg_td_oob chain
5 pairs -> 4 tiles, masks exact, avg/counts max|diff|=0 at tol 0;
test_convert_direct regression PASS both tiers.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

cuda_runtime.h only pulls device_launch_parameters.h under __CUDACC__,
which the indexer stopped defining when CDT 2025-09 dropped the Nsight
scanner-discovery provider -> 14 false markers (blockIdx/threadIdx/
__syncthreads/NAN/isnan unresolved; F3 still jumped via other TUs).
Fix: include cuda_runtime.h + device_launch_parameters.h + math.h
explicitly, plus a __CDT_PARSER__-only __syncthreads() decl (its real
declaration is __CUDACC__-gated). No-op for nvcc: standalone -dc compile
clean, build_tests OK, test_convert_direct PASS both tiers (strict 0).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- launches/test_convert_direct.launch.tmpl: harvested working shared config
  (C/C++ Application + cuda-gdb 12.8; .tmpl suffix so Eclipse does not
  discover the template copy as a duplicate; setup.sh strips it on install)
- setup.sh: install launches/
- nsight_per_kernel_tests.md: modern Eclipse does not load the Nsight local
  'CUDA GDB Application' type - use C/C++ Application with cuda-gdb as GDB;
  shared-config save recipe; DP child-launch stepping caveat
- README: document launches/ templating
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CLAUDE: gitignore live /launches (shared Eclipse debug configs; template copy will live in eclipse_setup/)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CLAUDE: eclipse_setup + build_tests.sh: ${HOME}-relative cuda-samples path (no absolute /home/elphel)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- .cproject: exclude src/tests|jna|attic from the managed build (tests are
  many main() executables built by build_tests.sh, not the one-binary
  managed project); add ${workspace_loc}/src include path (indexer resolves
  test #includes); sync Release samples include path with live drift
- .project: second builder = .externalToolBuilders/build_tests.launch
  (runs src/tests/build_tests.sh on manual Build; machine-generic via
  workspace_loc variables)
- .settings: adopt live doxygen + UTF-8 encoding prefs
- .gitignore: ignore live /.externalToolBuilders (template copies up, then
  disconnects, per convention)
- README.md: the convention + what is workspace-level (CUDA toolkit pref,
  12.8 cuda-gdb debug configs)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- make_convert_direct_case.py packages the legacy clt/ exports + goldens
  into a manifest case (one-time bridge; new cases come from Java oracle)
- test_convert_direct.cu: manifest -> RAII buffers (GpuBuf/GpuPitchedBuf +
  device pointer arrays) -> host-built tasks from portsxy -> single
  deterministic launch; the historical (0, kernels_hor) deconvolution-gate
  quirk is an explicit apply_kernels manifest parameter
- two-tier goldens: expected_clt_2025 legacy @ legacy_tol=1.2 (measured
  1.107 max, 2.2e-5 relative rms - same class JNA Stage2 accepted <1.0);
  expected_clt blessed on this machine @ tol=0 for real regression
- TpTestData.compareTo(named reference); avg_td re-verified PASS tol=0
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- src/tests/: TpTestData (manifest.txt contract with the Java oracle,
  testdata_format.md), test_avg_td main, build_tests.sh (CUDA 12.8,
  -arch=native, device-debug for NSight on Blackwell sm_120)
- src/tp_consolidate.cu: clt_average_sensors v0, bit-exact mirror of
  CuasTD.consolidateSensorsTD (NaN-absent, count plane, n>1 whole-tile
  poison, reciprocal-multiply op order); OOB + median/MAD stages next
- verified: synthetic + Java-oracle cases PASS at tol=0 on RTX 5060 Ti
- pre-existing edits in jna/tp_jna.cpp, src/TileProcessor.cu left alone
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CLAUDE: self-documenting comments: MCLT overview, quad algebra conj/no-conj, tp_task bridge, GPU geometry chain

Comment-only. From Andrey's explanation: MCLT perfect-reconstruction overview
(TileProcessor.h, with blog ref); convolveTiles = quad product WITHOUT
conjugation, kernels DTT-II/III family vs type-IV image tiles;
correlateAccumulateTiles = conj(T1) x T2, order-sensitive; shiftTileHor/Vert
= lossless fractional-shift phase rotators (differential rectification);
struct tp_task bridge role + composed-offset semantics; calc_rot_deriv /
calculate_tiles_offsets = GPU side of historic getPortsCoordinates().
Pre-existing dirty jna/tp_jna.cpp intentionally NOT included.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

tpdnn_infer processed the whole `count`-scene request in one shot; at full res
(512x640) a 64-scene field tensor is ~10GB (64x124x512x640) + decode p ~10GB ->
CUDA OOM on a 16GB card. The 64x80 synthetic smoke didn't expose it.

Now loops in GPU_CHUNK-sized sub-batches (== infer_server's GPU_CHUNK), writing each
chunk straight to the host output buffers so GPU tensors stay bounded (~3-4GB/chunk
at CHUNK=8). L2 hidden/age/sprev carry across chunks; reset only at the first chunk
when l2_reset (matches the server). Env TPDNN_GPU_CHUNK (default 8).

Validated: parity vs server oracle still EXACT (0.0) at CHUNK=8 (8+4) and CHUNK=4
(4+4+4) - L2 carry across chunk boundaries is correct.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Server-side copies the already-uploaded libtorch zip into
mirror.elphel.com/maven-dependencies/org/pytorch/libtorch-cxx11-cu128/2.7.1/ (maven
layout: .zip + .pom + .sha1/.md5, no 3.8GB re-upload) so imagej-elphel's -Plibtorch
profile can dependency:unpack it. Pairs with fetch_libtorch.sh (direct download).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Piece 2 of the native-JNA DNN path. tp_dnn.cpp is a C-ABI port of infer_server.py's
hot path so the Java client can run L1+L2 in-process instead of over TCP:
  tpdnn_init/upload/infer/free  (+ num_levels/level_frames)
faithfully reproducing build_pyramid, the 16x shift-and-stitch full-res recovery,
decode (ghostbuster + velocity centroid), and the L2 ConvGRU recurrence + track-age.
Loads the TorchScript models from imagej_elphel_dnn (export_torchscript /
export_l2_torchscript). Disables the TorchScript JIT fuser at init (nvrtc element-wise
fusion fails on Blackwell; production wants no runtime nvrtc).

Validated: native vs the running Python server (same CUDA) max|diff| offset5=0,
roi=0 — bit-for-bit. (Oracle dump_ref.py + driver tpdnn_test.cpp, scratch.)

Built standalone via build_dnn.sh (g++ + libtorch 2.7.1+cu128, ABI=1), separate
from the nvcc-built libtileproc.so; fetch_libtorch.sh pulls the pinned libtorch.
Context unification + zero-copy kernel<->tensor sharing is a later step.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>