It's documented behavior rather than a bug, but it's easy to miss because the doc lives on a different class than the one people call.

The nasty part is what happens when the cap hits. Per the docs, the search / resets and builds a path from the start polygon to the polygon that was found closest to the target position so far. So you don't get an empty path or an error you get a plausible-looking path that ends short. That reads exactly like a disconnected island.

This looks pretty cool!

32 is a deliberate ceiling for this edition, not a limit of the technique. The design goal was set no compute no buffers one file shapes travel as uniform arrays, and every primitive adds a distance eval to every march step on every pixel, so a fully analytic field doesn't want hundreds anyway. The scaling tricks are exactly the ones you list bake everything static into a precomputed SDF texture and sample it as a single primitive. That's the natural direction for the node based edition if it grows to v2

Hadn't seen the liquids part of that SIGGRAPH talk thanks for the pointer, watching it.

Nice transition effect Thanks for sharing!

Fair suspicion in 2026. AI can write it! , which is exactly why I built the sandbox instead of leaving this as a wall of advice. The useful part is meant to be testable version specific AStarGrid2D behavior, reproduced gotchas, and an interactive grid you can poke at. If you see a specific mistake, vague claim, or Godot-version mismatch, point me to it and I’ll fix it.