Fix three critical OpenCL-convolution bugs

[paxcalpt]

Found during an audit of the website branch. All three verified; these also affect PR #106 (cl_conv), which carries the same OpenCL commit.

1. convolve3D_new is undefined — crashes for every user (no GPU needed)

image_convolution.py had:

convolved_intensity = convolve3D(intensity_voxel, kernel3D)      # correct result
convolved_intensity = convolve3D_new(intensity_voxel, kernel3D)  # NameError

convolve3D_new is defined nowhere in the repo, so the volume-convolution imaging path (generate_frames_volume_convolution) raised NameError for everyone, GPU or not. The first line already computes the result with the real convolve3D; the second was a WIP leftover. Removed it. Verified: the module now imports and convolve3D_new is no longer referenced.

2. Separable 1D kernels applied to swapped axes — wrong GPU results

The separable decomposition extracts kx = kernel[:,cy,cx] (varies along axis 0), ky = kernel[cz,:,cx] (axis 1), kz = kernel[cz,cy,:] (axis 2), but applied ky on axis=0 and kx on axis=1. Each line must be convolved along the axis it varies on. Verified against a numpy/scipy reproduction: the swap produces ~7.4% error for a non-symmetric PSF; the corrected mapping gives 0. Invisible for axis-0/1-symmetric PSFs, wrong otherwise. (The GPU path itself isn't runtime-tested here — pyopencl isn't installed in the audit env — so the numerical check used the equivalent numpy logic.)

3. .cl kernel not packaged — GPU path breaks on pip install

[tool.setuptools.package-data] globbed only "*.yaml", so _convolution.cl wasn't included in the wheel; _get_cl_code's assert os.path.exists(...) would fail on an installed copy. Added "*.cl". Verified: a freshly built wheel now contains vlab4mic/utils/transform/_convolution.cl.

Scope deliberately limited to the three critical correctness fixes. The remaining audit findings (per-call OpenCL context/recompile, dead fp64 check, single-device GPU-type check, debug print; and the mkdocs packaging items) are filed as issues.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR fixes correctness and packaging issues in the OpenCL-enabled 3D convolution path used by volume-convolution image generation.

Changes:

• Corrects separable OpenCL convolution axis application.
• Removes an undefined convolve3D_new call that caused runtime failures.
• Includes .cl OpenCL kernel files in packaged distributions.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/vlab4mic/utils/transform/image_convolution.py	Fixes OpenCL separable convolution axis order and removes undefined fallback call.
pyproject.toml	Adds .cl files to setuptools package data so OpenCL kernels are included in wheels.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.