Add chance to draw local constants in HypothesisProvider

[tybug]

I'll be back with tests tomorrow, but figured it was worth PRing this for review.

An alternative to passing the local constants to PrimitiveProvider is to store it in BuildContext. Downside: ConjectureData instance behavior is no longer isolated and depends on context; harder to test. Upside:PrimitiveProvider implementers don't need to adjust their interface to access the constants, they can just opt-in to the BuildContext.

That said, CrosshairProvider uses def __init__(self, *args, **kwargs), so this shouldn't actually be a breaking change for them.

[Zac-HD]

Looks pretty sensible to me! Only design change I'd make - instead of passing local_constants as an argument to the provider, just let the provider collect them from whatever helper function we provide at construction time. My intuition is that few third-party providers will want to use this; it's basically a hypothesis-and-hypofuzz internals thing.

[tybug]

crosshair failures (eg) from latest version. I added a crosshair version pin and corresponding task to update it

Looking into the other actually-relevant failures on this branch

[tybug]

Maybe we should split the probabilities for GLOBAL_CONSTANTS and local_constants, so we still have a good chance to draw weird floats and strings even in the presence of many many local constants.

[tybug]

I've tracked down the pareto test failure to a potentially very old latent bug (manual front removal bypassing listeners) masked by our with deterministic_PRNG. Refactoring the test here rejigs the prng and hides the failure again.

I think we should merge this knowing that the test is incorrect, and follow up in the next pull. I have a fix, but want to write a new test for it and make sure it's correct:

diff --git i/hypothesis-python/src/hypothesis/internal/conjecture/pareto.py w/hypothesis-python/src/hypothesis/internal/conjecture/pareto.py
index 73f3e6a53..c282fa36e 100644
--- i/hypothesis-python/src/hypothesis/internal/conjecture/pareto.py
+++ w/hypothesis-python/src/hypothesis/internal/conjecture/pareto.py
@@ -248,7 +248,7 @@ class ParetoFront:
                 i -= 1
 
             for v in to_remove:
-                self.__remove(v)
+                self._remove(v)
             return data in self.front
         finally:
             self.__pending = None
@@ -272,7 +272,7 @@ class ParetoFront:
     def __len__(self) -> int:
         return len(self.front)
 
-    def __remove(self, data: ConjectureResult) -> None:
+    def _remove(self, data: ConjectureResult) -> None:
         try:
             self.front.remove(data)
         except ValueError:
@@ -338,7 +338,7 @@ class ParetoOptimiser:
                     # the front, or it was not added to it because it was
                     # dominated by something in it.
                     try:
-                        self.front.front.remove(source)
+                        self.front._remove(source)
                     except ValueError:
                         pass
                     return True

[Zac-HD]

(one comment, will have more later in the week)

[Zac-HD]

Minor comments considering the sheer size of this PR. I eagerly anticipate the many new bugs this will uncover downstream... merge when addressed to your satisfaction.

[Zac-HD]

I have a bad feeling about these pins. Can we just list hypothesis-crosshair (no pin) and -r test.in, and then install from crosshair.txt without test.txt?

Also happy to merge this PR and follow up later.

[tybug]

I'll follow up with this tomorrow, if only to unblock CI for other PRs!

[tybug]

I'm happy with this PR from my end, but I'm about to sleep and there's no need to rush things. I'll merge tomorrow, or you're welcome to merge as well :)