Support annotated_types in st.from_type

[Viicos]

Fixes #3356.

This is very much work in progress for now

[Zac-HD]

I think we should carefully avoid changing how strategies are inferred, and only use filters for this. Otherwise, we'll break the ability to register custom strategies for any of the types we expect might be annotated, such as datetimes! Implementation-wise, we'd limit our changes to:

hypothesis/hypothesis-python/src/hypothesis/strategies/_internal/types.py

Lines 368 to 369 in 058420a

	annotated_type = args[0]
	return st.from_type(annotated_type)

such that the only new code we'd have looks something like:

annotated_type, *constraints = args 
strategy = st.from_type(annotated_type)
for pred in constraints_to_filter_predicates:
    strategy = strategy.filter(pred)  # rewriting usually makes this efficient
return strategy

The contribution of this PR would thus be to implement constraints_to_filter_predicates() (the CONSTRAINTS_FILTER_MAP is great!), and to write tests which check that we only generate valid inputs from constrained types.

Once merged, I'll write up an issue for the feature-rewriting we'll want to add to make e.g. length bounds more efficient - the trick is that we can convert the constraints to callable objects which can also be recognised by the various filter methods. I think we can also rewrite length-constraining lambdas, so long as we keep the predicate just in case len was redefined... but that's for later.

[Viicos]

This sounds reasonable, the only issue is having timezones as a filter to times()/datetimes() can't be done without giving a timezones strategy beforehand (otherwise we would only get naive times/datetimes by default and the filter predicate would never succeed) :/

As you can see what I do currently is only special case times and datetimes to get the base strategy, otherwise I switch to from_type.

[Zac-HD]

Ouch, yeah. Let's dodge that by just not supporting Timezone metadata in this PR, and maybe emitting a warning 😅

I'm keen to work it out eventually, but imo it's much better to ship the parts we won't regret asap and spend as long as we need on the others.

[Viicos]

Let me know if the actual implementation suits you, I'll then add tests

[Zac-HD]

Looks reasonable! I'll review again once we've got tests and a changelog entry so that CI passes 😁

[Viicos]

(Still need to figure out how to test all constraints)

We probably need to tweak CI to test with annotated_types installed. Should I create an extra tox env?

[Viicos]

Thanks for finishing the implementation @Zac-HD! I'll try to think about solutions to:

• make filter rewriting more efficient (e.g. for len constraints)
• see if we can support timezones