nice - thanks, isn't using longs enough? that round with a hard-coded precision of 10 just looks weird.
best approach would be to detect long overflows on num/del and simply handle those in a different branch.

@kares Thanks for your input. I agree, the rounding part is a subpar solution.

However, while using longs fixes the problem with addition, it's not enough to fix the problems with subtraction. For instance, within the provided spec, the result for val in op_plus_numeric will be -864.0000000000001. The imprecision of the value will flow through the whole calculation causing odd effects (e.g. (dt - (dt - 0.0001)).to_f => 0.0001 but (dt - (dt - 0.00001)).to_f => -0.9999899884258123.

I'm not quite yet there to understand how and where to best deal with the problem though.

yeah having floats around is always problematic, think I have some ideas on how to proceed.
we should be able to somehow detect the "over-flowing" sub-traction case you mentioned.

also as a work-around you should be able to run fine with JRuby 9.1 .. 9.2.1 release should be close thought

for the record, only the add part is to be considered a regression, the sub-tract doesn't work in 9.1 either:

DateTime#- is able to subtract sub-millisecond precision values FAILED
Expected (7599824371187713/759982437118771200000)
 to equal (1/100000)

... we do the calculation "right" in 9.2 ~ the same as in previous JRubies.

can not use the rounding as is since (as CI revealed) that introduces more problems, will try to port the Date#-calculation logic to semi-match MRI, its tricky since we're using JODA while MRI has raw structs.

pushed a modified commit (with the long fix since its a regression) into master: bf2a83c ... so this doesn't hold up for long. tagged the failing subtract specs, might push a fix for that later - have one but it doesn't scale beyond 5 decimal places.

going to close this one, the provided specs are now passing ... there's still more work to do.
guess we'll track forward as users report, unless some of use decides to pro-actively work on it :)