Fixes two places where we relied on map iteration order when processing
values, which leaked nondeterminism into the generated SAT formulas.
Adds a couple of tests that directly assert that the SAT system is
equivalent on each run.
It's desirable that the formulas are deterministic based on the input:
- our SAT solver is naive and perfermance is sensitive to even simple semantics-preserving transformations like A|B to B|A. (e.g. it's likely to choose a different variable to split on). Timeout failures are bad, but *flaky* ones are terrible to debug.
- similarly when debugging, it's important to have a consistent understanding of what e.g. "V23" means across runs.
Both changes in this patch were isolated from a nullability analysis of
real-world code which was extremely slow, spending ages in the SAT
solver at "random" points that varied on each run.
I've included a reduced version of the code as a regression test.
One of the changes shows up directly as flow-condition nondeterminism
with a no-op analysis, the other relied on bits of the nullability
analysis but I found a synthetic example to show the problem.