This bug catcher is so good that it successfully reported 2 bugs on the first attempt.

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 88.38%. Comparing base (c6a79f6) to head (a410361).
Report is 32 commits behind head on master.

@@            Coverage Diff             @@
##           master    #6878      +/-   ##
==========================================
- Coverage   88.40%   88.38%   -0.02%     
==========================================
  Files         326      329       +3     
  Lines       71946    72468     +522     
==========================================
+ Hits        63605    64053     +448     
- Misses       8341     8415      +74     

I'm so proud of my dilligent work. :)

However, as I unintentionally wrote "This does not solve any bugs, but reports them earlier.". :)

• Now I reimplemented the code that creates labels for discretized variables so that it increases the number of digits when necessary. When the number of digits is given as an argument (as opposed to using ContinuousVariable.str_val), it also rounds the interval threshold so that they match the printed labels.
• The other bug was due to an invalid test.

Ok, I found some nitpicky issues. I can fix it myself if you agree with my suggestions.

I replaced np.round with round. The results are a bit different, but no implementation was wrong (see tests).

I could not make a test case up where the old implementation would have actual problems, for example, points that would not correspond exactly to limits, but I can not guarantee one does not exist either. So, this seems safer, but I can not really say that the previous implementation was wrong.

The difference between Python's and numpy's rounding.

n = 2.3455; npn = np.float_(n)
round(n, 3)
Out[17]: 2.345
round(npn, 3)  # this one is fun
Out[18]: 2.346
print("%0.3f" % n)
2.345
print("%0.3f" % npn)  # :)
2.345