To help people reviewing and prevent them from needing to closely study the code to get a mental model of what happens I'll write a short outline of the code path here.

All functions reside in the file source/blender/editors/sculpt_paint/paint_image_proj.c

When projecting an image from the camera onto the mesh

texture_paint_camera_project_exec(...)

is called. This does some setup and builds up a PaintProjectState object (ps) , attaching (among other stuff) the reproject_ibuf (the source image) and the projImages array (the destination images). It then calls

project_paint_op(...)

this sets up a task_pool, does some more setup and fires up the taskpool with the function
do_projectpaint_thread(...)
as threadfunction.
Inside do_projectpaint_thread(...) in the branch where ps->source is not PROJ_SRC_VIEW and not PROJ_SRC_VIEW_FILL it does a check if the ps->reproject_ibuf and ps->projImages[currentlyactiveimage] both have a rect (uchar data) or a rect_float (float data) and if not create the missing data in the ps->reproj_ibuf. But since this is run in a multithreaded context, while the ps->reproj_ibuf is a shared object, will lead to a race condition, both in the check and in the generating of the missing ibuf.rect(_float) .

Now that I've written this down I think this will go wrong any time you do projection painting with float textures on a material containing uchar textures or vice versa. So probably also if you use a float texture brush on a material with uchar textures and v.v.

I hope this writeup makes it easier to understand what I'm trying to do with my patch. And also easier for someone who actually knows the code to maybe create a better patch,

Patch LGTM, besides nit-picky style comments below.