I've had this idea for a while—we should use binary search to compute an optimal precision allocation that we can compare Rival to. That code is in this PR; it's AI-written but I did manually refactor everything. There's also an infra/optimize.rkt script that does the optimal precision allocation for every single point in the points.json file, in an attempt to find the ones that can be improved.

So far, looking at the results, I've learned:

• In quite a number of cases, adjusting after the final iteration doesn't work; I wonder why, we should fix this
• Our initial precision allocations are often too high; I don't think that's super interesting so I'd like to filter it out
• When the output underflows, Rival attempts to compute ~53 bits of the output but it actually only needs to bound the exponent, and in this case over-estimates

Excited to see what else happens!