Annotate SearchStrategy.filter with TypeGuard overload

[CharString]

TypeIs[T] case is not expressible (yet?), because you can't express the intersection type:

@overload
def filter(self, condition: Callable[[Ex], TypeIs[T]]) -> "SearchStrategy[Ex & T]": ...

[Liam-DeVoe]

I don't understand the motivation behind this change. TypeGuard is for use in narrowing different branches of a conditional, and I don't see how that gets us anything beyond the current annotations. Can you give an example of a before/after type check that is improved by this change?

[CharString]

@Liam-DeVoe A filter is a different branch of a conditional, in the same sense tail recursion is a for loop.

For some strategies Ex is very broad. e.g. hypothesis_jsonschema.from_schema: SearchStrategy[JSONType] a TypeGuard can narrow it down.

The semantics from set theoretic types (i.e. types are sets of values) are that a strategy samples from the Ex set and that filter defines a subset of Ex. And true, any sample from the filter is still also in Ex, but often that subset has a name too. TypeGuard allows us to express that name.

And in a world where a bunch of code is generated by agents, having these types explicit and readable in our PBTs, helps with raising trust in the code.

[Liam-DeVoe]

Here's an example of strategy whose inferred type changes:

from typing import TypeGuard

def is_str(x: object) -> TypeGuard[str]:
    return isinstance(x, str)

s = from_type(object).filter(is_str)
# previously: SearchStrategy[object]
# this PR: SearchStrategy[str]
reveal_type(s)

In the future it would help to give a motivating example in PRs like this!

Let's add some tests to revealed_types.py that use type guards in .filter.