I’ve solved the problem! At least for myself, hopefully there will be other lucky people.
After reading the Snapmik’s comment, when they’ve realised there’s a difference in the linear module speeds, I’ve tried the same and got the same result. And indeed it comes back to the same state when you go back along Y axis, so it’s a different steps/mm for the modules.
I’ve decided to check how other linear modules behave. I had to disassemble the printer completely.
And make a bunch of primitive tests. Put the modules side by side in pairs, connect both to Y axis, and try the same experiment. It’s almost like horse race, you can even bet on them!
That’s how it looks on the max Y (sort of 350mm on A350 I guess) after perfectly aligning the marks at the start.
In the end I found 3 modules with exactly the same speed. Two others are around 1 mm “shorter”, and nearly the same. Not sure if truely the same, but my experiment precision is not great.
Anyway, I can put two “shorter” once to Z axis (since it’s very long distance between them, I guess it can tolerate a slight difference, even if it exists). Three others would make X and Y.
Also I’ve noticed I had some issues during CNC, when calibrating the head perfectly on the left corner, there’s ~1mm gap on the right side with the same programmatic height. I guess that’s where my other “short” module used to be, so it should be fixed now too!
I haven’t tested the result yet. Moving home, so I had a good opportunity to disassemble the printer, will assemble back in 2 weeks to try. But I have a really good feeling!
Hope it helps somebody.