[endrix]

A medium-level review of #112, #114 and #115 after their merge. Six of its ten findings held up and are fixed here, each with a test where one could be written:

• The drop still swapped routes. Once the server's route reached the ports, the view fell through to the live-router branch for the rest of the 650 ms settling window, then jumped onto the server's route. The server's route is now drawn the moment it reaches the ports.
• The reveal fallback was flat at 1.5 s while libavoid's time grows faster than the edge count. It now waits 1.5 s plus 25 ms per hidden edge, 6 s at most.
• A node with no edges was walked on every pointer move. Once per drag now.
• The trace turned a wire between two overlapping nodes because it compared port centres. It compares node centres now; a group's boundary port keeps its own place.
• A slid route could keep a 1 px stub (a node moved up to its own bend), a vertical exit to the eye. A kept route needs an 8 px stub, else it is routed anew.
• unkink guessed which end it was given by x-distance and picked the wrong one on a U-shaped route. It takes the end explicitly; libavoid lists points source to target. The unpadded node boxes are derived from the padded obstacles instead of a second model walk.

Left as they are, with the reasons: a pointer released outside the diagram sends no finished move, so its edges stay hidden up to the 5 s abandoned-drag timer (the price of keeping a node held still mid-drag hidden); the bootstrap reroute of a diagram with a saved layout and no saved routes goes through libavoid for up to 200 edges, a one-time cost at open that the perf line reports; the hidden-edge state stays a module map the view polls, as the other drag state is.

https://claude.ai/code/session_01EToj6RJRMSSHQiqGNMU9wv