It's using a set purely keyed by number of remaining bytes, so if there are two pages with the exact same number of remaining bytes, one of them gets lost.
See #1077
The color picker returns color values up to 255.99 so this can otherwise result in it returning a color like #100100100, and the hex code otherwise not matching with what is displayed in the color picker.
In the 1.10.8 thirdparty packages, OpenCV is split into a lib folder and a lib-10.6 folder, so we should use the proper dylibs to still be able to target 10.6
Besides use of --universal with --osxtarget 11, it's now also possible to create a fat build with ARM64 by using --osxtarget 10.9 --arch x86_64 --arch arm64 (it will use the 11.x SDK but target 10.9)
FMOD Ex is also being disabled when building against the 10.14 SDK or higher, because the necessary libstdc++.6.0.9.dylib is no longer being shipped. FMOD Ex, libRocket and Cg are all disabled on ARM64 (FMOD Ex and Cg don't have the binaries available, libRocket could theoretically be built but there's no point as there won't be an ARM64 build of Python 2.7, and besides, we've deprecated it anyway).
Fixes#1062
When a LightAttrib is set on the root of a model and the light is contained within that model, creating a copy (like the one Loader makes automatically) causes the LightAttrib to still point to the original light, not to the newly copied light.
This works around it only for the case when the LightAttrib is set on the root, and only for the case of NodePath::copy_to() (it turns out there are edge cases when doing it in PandaNode::copy_subgraph() that would be hard to explain).
See Moguri/blend2bam#44
Panda was adding a column with weights (0, 0, 0, 0), but these weights don't add up to 1 so no useful identity matrix can be produced. Instead it's better to fall back to the OpenGL default, (0, 0, 0, 1).
I'm also defaulting the transform_index values to (0, 1, 2, 3) to support non-indexed skinning (although that's pretty esoteric, given that that only supports 4 transforms...)
This is necessary to prevent deadlocking on the GIL if some Python code is trying to write some geometry while the CollisionTraverser is trying to read it.
Fixes#1033
