Thanks for giving it a go!
The second-to-last level relies on an asymmetry in how dogs respond to seeing that their target is no longer in the location they expect. For treats (which don't usually run away under normal circumstances) they just give up and stop wherever they are, as opposed to going all the way to last known location (as when chasing the player). This really only happens once before in all the levels, so the punchline here definitely needs better setup in the preceding levels.
The last level demands a very precise sequence of biscuit collection while also allowing the player to make some irrevocable mistakes early on. While I don't think this is inherently a bad thing, the game as-is forces a lot time to be spent executing individual moves even if the player has a very clear global sequence in mind. It would be interesting to explore ways to make movement planning and iteration go quicker.
As far as dog-stacking goes, it felt like the more natural way to go. With a little more time to polish, I would definitely want to implement some small offsets for each stacked dog to give the illusion of a pack all running together. Alas, this ultimately had too many mini problems to solve to make it a priority for the jam submission.