Extend run_convert_pipeline with do_corr: after convert_direct, allocate corr buffers
(gpu_corrs_td/combo_td/combo via alloc_image_gpu + corr_indices/combo_indices/num_corr_tiles),
launch correlate2D <<<1,1>>> (TD, CDP; 18 args, generates indices), read num_corr_tiles,
corr2D_combine (quad pairs_mask 0x0f), corr2D_normalize (TD->pixel), de-pitch gpu_corrs_combo.
tp_corr_selftest wrapper (do_corr=1).

Validated on RTX 5060 Ti via JNA: num_pairs=120, num_corr_combo=5120, output stats identical
to golden (max 0.6638, rms 0.0717). clt/aux_corr-quad.corr is OLDER (Apr-2025) than the CLT
golden (Jul-2025) so the active-tile ORDER differs -> pointwise compare is permutation-dominated
(0.66). Order-independent check (sort both, compare distributions): max value error 2.06e-05
== float32 precision => correlate2D/combine/normalize compute the correct values on Blackwell.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Refactor the Stage-2 selftest into run_convert_pipeline(do_imclt) shared helper;
tp_convert_direct_selftest is now a thin wrapper (do_imclt=0). Add tp_imclt_selftest
(do_imclt=1): after convert_direct, allocate pitched RBG output buffers (alloc_image_gpu,
648x520/cam, mono), launch imclt_rbg_all <<<1,1>>> (gpu_clt -> gpu_corr_images),
de-pitch via cudaMemcpy2D, compare to clt/aux_chnN.rbg golden.

Validated on RTX 5060 Ti via Java->JNA: max|RBG-golden|=0.0201 over peaks 1535 ->
relative ~1.31e-5. convert_direct CLT error unchanged (0.108505) => no regression.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add tp_convert_direct_selftest to the JNA shim: mirrors TpHostGpu allTests' convert
path (setImageKernels/setImgBuffers/setCltBuffers/setTasks + calc_reverse_distortions
-> rot_derivs -> calculate_tiles_offsets [CDP] -> convert_direct), reusing the harness
runtime-API host helpers (tp_utils/tp_files/TpParams/tp_paths) for ALL allocation and
porting only the launches to driver-API cuLaunchKernel against the NVRTC module. Reads
CLT back, compares to clt/aux_chnN.clt golden.

build_lib.sh: nvcc + -std=c++17 (static constexpr TpParams members become inline),
-Isrc + cuda-samples Common (helper_cuda.h), --pre-include algorithm.

Validated on RTX 5060 Ti via Java->JNA: num_active_tiles=5120 (all), max|CLT-golden|
=0.1085 over peaks of 12260 -> relative ~8.85e-6 (float32 NVRTC-vs-nvcc variation).
First CDP (calculate_tiles_offsets) and 17-arg pointer-of-pointers convert_direct
launch executing natively on Blackwell.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add TpInstance to the JNA shim: device buffers (gpu_geometry_correction,
gpu_rByRDist, gpu_rot_deriv, gpu_correction_vector) + setters (HtoD),
the two pure-geometry launches (calcReverseDistortionTable {16,1,1}/{3,3,3},
calc_rot_deriv {num_cams,1,1}/{3,3,3}), and readback getters. Driver-API
cuLaunchKernel against the NVRTC module (mirrors GpuQuad.execCalcReverseDistortions
/ execRotDerivs, no JCuda). build_lib.sh builds libtileproc.so.

Validated via Java->JNA against tile_processor_gpu/clt reference data on the
RTX 5060 Ti: rByRDist == clt/*.rbyrdist to ~1e-7 (aux 16-cam and main),
rot_deriv rows orthogonal to ~1e-10 (scaled-rotation structure, det~zoom^3).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

libtileproc shim (tp_jna.cpp: extern "C" tp_create_module/num_functions/last_error/destroy)
+ standalone tp_nvrtc_probe.cpp + build_probe.sh. NVRTC-compiles the kernels (+ JCUDA defines)
-> cuLink(libcudadevrt, CDP) -> module -> 19 functions, validated on the RTX 5060 Ti (sm_120 via
compute_90 PTX + driver JIT). Build artifacts gitignored. By the JCuda->JNA migration.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>