Fix out-of-bounds read in wavelet decomposition on small images (e.g., thumbnails)

[masterpiga]

Problem

The à trous wavelet decomposition in common/dwt.c averages each pixel with neighbours hscale pixels to the left/right (and rows above/below), where hscale doubles each scale level. At the borders, out-of-range neighbours are reflected back inside using 2*width-2-(col+hscale).

The horizontal pass capped the scale at MIN(1 << lev, width). When hscale reaches width — a small image (e.g. thumbnail/mipmap export) combined with a high wavelet level — the reflection of the last column underflows to -1. Since the index is an unsigned size_t, it wraps to a huge value and the code reads gigabytes past the buffer → SIGSEGV.

This reliably crashed retouch during thumbnail export (_image_load_job_run → pixelpipe → retouch → dwt_decompose). The vertical pass already capped correctly at height-1; the horizontal pass did not.

Fix

• dwt_decompose_horiz: cap hscale at width-1, mirroring the vertical pass.
• dwt_denoise_vert_1ch / dwt_denoise_horiz_1ch: same size-1 cap. The horizontal denoise kernel additionally used three disjoint loops that overlap and double-process columns when 2*hscale > width; reworked the loop bounds into a true non-overlapping partition and reflected the far neighbour in the edge loop. The hot middle loop keeps its vectorized form.
• OpenCL (dwt.cl + host): clamp sc at size-1 and reflect the far neighbour in the left/top kernel branches, matching the CPU path.

Behaviour is unchanged in the valid range (hscale < width); only the previously-crashing over-scaled case now produces a well-defined max-reflection result.

macOS crash report

crash_report.txt

Co-authored with Claude.

[masterpiga]

@TurboGit also in this case, I tentatively set 5.6.1 as the milestone. The bug is not 5.6-related, but the change is a genuine bugfix.

[TurboGit]

Thanks! I'll merge on master for now and if regression tests are clean tomorrow and my testing later today is good I'll merge for 5.6.1.

[jenshannoschwalm]

Just checked integration tests also for OpenCL with just-now commits. All good.