Most cells are accessible, so presizing to the full size of the grid is sensible to avoid allocations to resize as it is filled up. The set can also be reused across all layers to avoid allocating it many times.
When a search is initiated from an inaccessible source location, a path is still allowed if there is an adjacent, accessible location the unit can move into. The local pathfinder and HierarchicalPathFinder already account for this logic, but HPF has some bugs.
Firstly, when the HierarchicalPathFinder heuristic is being evaluated, it assumes all cells being explored are accessible - this is important for performance as it avoids constantly rechecking the accessibility of cells. Although this fact holds true for cells explored by the path search, it does not hold true for cells being added as the initial starting points of the search.
Secondly, when checking for adjacent locations to an inaccessible source cell, we checked only if these were still on the map. This is insufficient - we need to know if movement between the source cell and the adjacent cell is possible.
The fixes resolve this by:
- Teaching the heuristic an extra parameter to know if the location is known to be accessible. This allow an accessibility check to be performed for starting locations which stops HPF mistakenly assuming the abstract node for that location is the one we need to consider, and to correctly check the adjacent locations and their abstract nodes instead. The parameter means will can still skip the accessibility check in the typical case where the path search is being expanded, preserving performance.
- When adjacent cells are considered we now check if movement to them is possible. This stops HPF from allowing jumps over height discontinuities (i.e. no magically jumping up or down cliffs) and thinking a path is possible when it is not.
