st.floats() with bounds do not shrink correctly

[Zac-HD]

The floats() strategy has a well-defined shrinking order: the fractional part is simplified away, then it shrinks towards zero as for integers(). While this is the case for unbounded floats, floats with bounds usually do not follow this shrinking order:

hypothesis/hypothesis-python/src/hypothesis/searchstrategy/numbers.py

Lines 185 to 188 in 7d1cf43

	def do_draw(self, data):
	f = self.lower_bound + (
	self.upper_bound - self.lower_bound
	) * d.fractional_float(data)

Instead, they shrink by simplifying the fraction of the allowed interval, with the lower bound as the minimal value. This technique is also lossy for large ranges, as it can generate at most 2^52 distinct values. The way that floats with one bound shrink as an offset to that bound is a relatively minor problem.

To fix this in the general case we will probably have to use the "shrink open" trick explained here and demonstrated here (from #1621), to split generation from shrinking.

[Zac-HD]

I was imagining [...] a single wrapper strategy called say FloatShrinkingStrategy which wraps some underlying strategy and filter function (or set of float specific constraints) and automatically produces a float strategy that shrinks correctly. This would then commonize shrinking logic between all three float strategies, and potentially would allow us to (later) improve the generation of floats to boot.

@DRMacIver, in #1644 (comment)

[Zac-HD]

Closing in favor of 2907.