Liminalia fixes the two bugs the non-Euclidean research pass exposed

liminalia

A research pass into how other non-Euclidean games handle space — Antichamber, Manifold Garden, HyperRogue, Hyperbolica, wrapped-world literature — confirmed the graph-first design dodges the seam-handling problem class entirely. It also exposed two real bugs in citizen movement and one open design question. This pass closes all three.

edge-exact routes

RouteResult stored node ids only, so movement re-derived edges from node pairs and traversed whichever parallel edge appeared first in adjacency order. A road and a spatial fold between the same two nodes could be priced through one and walked through the other — emitting spurious fold crossings and wrong travel times. RouteResult now carries Edges parallel to Nodes (edges[i] connects nodes[i] to nodes[i+1]), Dijkstra tracks the previous edge, and movement traverses by edge id.

re-planning on edge removal

Removing an edge mid-route froze citizens: UpdateMovement broke on the missing edge without clearing the route, leaving IsTravelling true forever and blocking every goal gated on !IsTravelling. Movement now re-plans from the citizen's current node when the next route edge is gone, abandoning the trip cleanly when the destination became unreachable.

return-safe routing

The design question: one-way folds are a legitimate non-Euclidean trick, but nothing stopped a citizen routing through one into a region they could never leave. Decision — a one-way should never strand. Routing is now anchored: a citizen's route may only pass through nodes from which their home node is reachable, computed by a new reverse traversal (TopologyGraph.Predecessors) that respects edge direction. A one-way fold into a pocket home can't be reached from is never routed through, even when it's the cheapest path. A fold pair (one-way out, one-way back) stays a valid shortcut. Citizens already outside the safe region fall back to unconstrained routing so they're never frozen in place.

The verification pass caught two of my own mistakes: the first regression test didn't pin the bug (the dead-end was off the path, so Dijkstra never went near it), and the graph's adjacency only registered one-way edges at their source node, so reverse traversal couldn't see them at the destination. Both fixed — every edge now registers at both endpoints with direction filtered at iteration time.

Both regression tests verified to fail on the pre-fix code. 215 tests passing.