Hans Hagen
After a day of testing, tracing, wondering etc Luigi and I figured out what was the problem it tooks us lots of debug prints -). Indeed there is a bug in luatex.
Great detective work!
So, some random value (not entirely random as it depends on what mem area the glue node is allocated from and that is kind of stable in a run) was added which results in the effects you noticed.
Ah, so running it 30000 times would allocate the glue node from higher and higher mem, and thereby making the not-so-random value larger and larger.
So, once this fix moves upstream you can check your suspicious cases. Luigi will push it in our experimental branch and I suppose that Taco will then push it into trunk and after that you can compile
Once it in the luatex trunk, I will check it right away. I am also happy to test the experimental branch if you give me a pointer to it.
or wait till the garden carries it.
Do you mean in the minimals?
Thanks for noticing!
For sure. I am very happy that it is fixed, and fixed before I have to give in my absolute final PDF file. My only worry is that the carefully tuned page breaks, some of which are sensitive to 1pt insertions of vertical space, might change once the underbraces return to their homes. But I'll tweak it with a bit of vertical space. -Sanjoy