The `assume(init_items != init_unequal)` was not sufficient
because it would allow cases like:
bi_cls=bidict, other_cls=bidict, init_items=[(1, 0), (0, 1)], init_unequal=[(0, 1), (1, 0)]
causing the resulting bidicts' item sets to still be equal,
so assertions like
assert items_match(some_bidict, other_unequal, relation=ne)
would fail.
Changing the assume to
assume(set(init_items) != set(init_unequal))
fixes the issue.
Surprisingly, hypothesis only just found a failing example for this
for the first time (in https://travis-ci.org/jab/bidict/jobs/363962257)
despite having numerous opportunities to find it before.
Using a new max_examples=5000 hypothesis profile, only on nightly CI,
should help surface any remaining issues that hypothesis hasn't found yet,
without slowing commit-triggered tests with too much example generation.
Also improve Travis config:
- ditch ALLOW_FAIL=1, for some reason it didn't work
- comment out python 3.7-dev job while Travis is stuck on an old 3.7 alpha
- don't use "- |" command block where we want to see output after each command
- prefix more network-dependent commands with travis_retry
Finally, improve docs building setup:
- Factor out sphinx from "dev" extras_require to its own "docs" extras_require.
- Add a .readthedocs.yml to customize the readthedocs build.
Use pip and the "docs" extras_require to get an up-to-date Sphinx version.
Fixes some broken :ref:`addendum:...` autosectionlabel links in learning-from-bidict.
jab