We discard what we've "learned" after an hour. We could degrade faster, or we could try to measure channels recovery speed. This requires more analysis though!
Your data would be a useful starting point!
Dunno if you saw https://bluematt.bitcoin.ninja/2024/11/22/ln-routing-replay/ but I recently started being more rigorous about our pathfinding scorer. Might be something to play with, it seems like the simple “just keep upper and lower bound on each channel’s liquidity” approach performs *worse* than always assigning each hop a 50% success probability. Keeping a histogram of those bounds, though, does reasonably good.
Discussion
At least my initial analysis of the data seems to say that degrading instantly is better than any other time constant 😭. (May just be some nasty bug in the way we’re calculating probabilities?)