LeanCheck property tests are biased towards the first argument

[robrix]

@tclem & I found a while back that a property test with four args would use the same value for the latter three on every iteration. It’s like how digits work: you enumerate all the digits in the rightmost place before incrementing the next one, and enumerating all the digits in the rightmost column again.

If you only do a thousand tests, we have enough Category constructors to almost exhaust that right there. Multiply that by the size of Syntax and so forth and you’d have to do an incredible number of tests before trying a different value for the second arg. We no longer have those types, but the problem still occurs when we have annotations which can take multiple values.

The enumeration we do right now is column-major, but we’d prefer an ordering that explores the space in a more balanced way; the Hilbert curve comes to mind, but Morton ordering may be more appropriate.

Alternatively, maybe we should move back to using QuickCheck or another stochastic property testing framework like hedgehog.