num-bigint VS advent-of-code

I'm trying div for real: https://github.com/rust-num/num-bigint/pull/236

Using an enum to simulate dynamic typing: https://github.com/philipc/prechelt-phone-number-encoding/tree/rust-number. This should be similar to the performance improvements from something like https://github.com/rust-num/num-bigint/issues/36, which means suboptimal use doesn't hurt as much.

edit: See for example https://github.com/rust-num/num-bigint/blob/7562ab24330792817e42b808f60b0cac51ca261a/src/macros.rs

So I wrote some code that can solve Part 1 of Day 23 in about 1.6 seconds for the sample input, and about 53 seconds for my actual input. Not great, but okay. It uses the A* implementation I originally wrote for Day 15, and then modularized for Day 23 (and hopefully future years). Then I updated the code for Part 2 (in two separate commits, this and this). These updates made it significantly worse. :/ Now, when I run Part 1 for the sample input, it takes about 10 seconds to find the answer. Earlier, I started it looking for the Part 2 answer on the sample input, then went and ate dinner and watched a TV show, and 45+ minutes later, it was still running and hadn't found the solution yet, so I killed it. Does anyone have any suggestions on what I might have done to make things so much worse, and/or how to optimize what I have now? I don't mind something that takes a couple of minutes potentially to find the solution, maybe even up to 10-15 minutes on a hefty problem, but this is going to take forever to find the solution.

As someone who has completed every year of Advent of Code so far, I read Part 1 and said to myself, "I'd better just go ahead and use a linked list for this; part 2's gonna be huge." My AoC project (in TypeScript) has my own custom doubly-linked list implementation that has served me pretty well. My runtime for part 2 of this challenge was ~14 seconds.