Weird thing: On Windows it renders but is slower. On Linux and Mac OS it does not go beyond Sample 1, it constantly refreshes.

Hi Lukas and Thomas.

First of all , a big thank you for having a look at that, much appreciated.

A detail that may have its importance (at least intuitively I'd say it may have…) is that in the particle system, I chosed to tick the "use modifier stack" option of the "emission" subpanel. As the behaviour is different when the stack is applied or not, I thought that may be of importance.

I will try to reproduce the error on another file and single our if that particular option is relevant in that case.

EDIT: no success. The test file I've made showed no particuliar impact from that specific toggle.

This is broken since the threaded despgraph update (rB709041ed0b7e1848068c9d53543ed114229a9f5b)

Seems to be caused by crappy particle check in the modifier:
https://developer.blender.org/diffusion/B/browse/master/source/blender/modifiers/intern/MOD_particlesystem.c;f8c247f21b8e489750e3847a4a2270c1146cbbc2$182

Somehow when the modifier stack was evaluated in 2.69 it would (imo correctly) use the viewport settings instead of render settings for the modifier stack. I'm actually not sure how this worked, since what Cycles does suggests it would always use the render settings!
https://developer.blender.org/diffusion/B/browse/master/intern/cycles/blender/blender_object.cpp;f8c247f21b8e489750e3847a4a2270c1146cbbc2$449
(1==eModifierMode_Realtime, 2==eModifierMode_Render)

This is the same code in 2.69 and master, not sure what prevented modifiers from using "realtime" settings before.

Further there is some kind of randomness in the skin modifier! For some reason it always generates slightly different vertex/face counts each time it is evaluated. This might be the reason the particle system subsequently updates over and over, because it detects a modified input mesh and re-distributes particles.