I wish we had better reproducibility tools. It should be possible to write a failing test case, and everything that is not of value to that test case gets stripped out, containing just the parts of the program/environment that can affect the test case (including sources of non-determinism like clocks and RNGs). Kinda like C-Reduce, but on steroids.

And we should be able to look at the problem probabilistically too: given the form of the test case, the value pipeline through the debugger, and the structure of the program, we should be able to narrow down places to look as much as possible using some form of bayesian search theory.