1. Why existing navmesh tools, such as Recast, fall short, and why adapting them requires more than a few tweaks. Recast’s walkable-slope limit can’t simply be disabled: the “which way is up” assumption is by design, and it’s baked into its heightfield data structures and pipeline stages. Adapting it is a rewrite, not a tweak, and understanding why tells you what a replacement must do differently.

2. A generation pipeline for navmeshes on any surface, and the one design decision that makes it work. Distance volume → iso-surface extraction at agent radius → quadric simplification → convex merging → navmesh construction. The key idea: extract the surface one agent-radius out from collision, so the navmesh tracks the agent’s centre. That one choice simultaneously makes every location valid at any orientation, removes the need for fit-filtering, and simplifies path following.

3. How an “unfolding” approach makes shortest-path finding work at any orientation. Standard 2D funnel algorithms fail when corridors twist through 3D space. However, flattening the corridor by hinging preserves its length, allowing the classic algorithm to run unmodified in the flattened plane before mapping the results directly back to the 3D surface.

4. The tradeoffs our implementation makes, and why. Most major design decisions brought robustness by giving something up: treating agents as spheres, keeping the distance field unsigned (accepting some unreachable interior mesh over a brittle inside/outside test on real geometry), keeping Detour’s corridor-first search design unchanged. Plus the unresolved challenges and gaps that remain.