Should I feel silly? I wondered how many "failing" items actually go to the real solver and the answer is 0. My challenge input has no "close calls", and pre-solver pruning check of grid area being enough for all required tiles solves the challenge in milliseconds instead of around ten minutes.

I still need to write a visualizer to better understand how it really works inside, it's that clever. And memory hungry, at its peak the solver used over 30GiB of active memory running a few in parallel. (3/3)

After a bit of research on coverage problems I ended up watching Donald Knuth explaining his Dancing Links algorithm for an hour and a half. Bit more reading and writing later I had a working solver and finally a solution. (2/3)

Sinällään jännää, miten hankalaa on ollut oppia suomennoksen sanat tuttuihin biiseihin joita on laulanut viroksi. Osansa on toki sillä, etteivät suomennoksen sanapainot istu aina luontevasti, koska sävellys on tehty todella hyvin viron sanapainoja hyödyntäen.

MathML tuki riittäisi suurimpaan osaan tarpeista. Monet wikimoottoritkin tukevat sitä.

Okay, had to give up on brute-forcing it. It was still only 1/4 done, so it would have taken ~34hrs.

Polyglot Notebook version of DFS with contracted edges took 12s, and AOT compiled C# native binary took 7s. Of course it used a dynamic programming cache to speed things up with a "small" factor.

ps. I did this in VScode Polyglot Notebook, and the visualization updates every 100k steps at around 2fps. Cyan is current status, green is longest known route thus far.