This was a fun one! My initial implementation, before I looked at the data, tried a similar technique as the one with the guard and the rocks - delete each wall square, rerun path calculation, note difference - which worked for part 1, but was stupidly slow (like 70 seconds). And then I saw part 2 and it didn’t work there anyway

So I thought about it a bit, looked at the data, noted that every single floor square was part of the path, so I didn’t need to do any path recalculation - I just needed to slice parts out of the calculated path. And I didn’t even need to do that, I could index each item in the path and work out the difference.

Part 2 is a wee bit slow because I worked out allllll the options within 20 squares of each path square and calculated the difference, but it’s still under a second!

(and trying to optimize with Task.async_stream made it an order of magnitude slower)

Name                     ips        average  deviation         median         99th %
day 20, part 1         13.45       74.33 ms     ±6.84%       74.14 ms       83.83 ms
day 20, part 2          1.30      771.43 ms     ±1.61%      767.13 ms      796.90 ms