Why the daily dungeon needs a fair shuffle

Every day the game rebuilds the dungeon from the designer's plan. Each room keeps its number of doors: the boss room always has one, the town square has many. Only which rooms the corridors join changes. "Fair" means every possible wiring is equally likely, so regulars can't learn the plan. The quick shuffle most games use keeps too many of the designer's own corridors. Dealing doors out at random and redealing whenever two corridors clash is perfectly fair and very fast for ordinary rooms, but with several big hubs almost every deal clashes. The new method trades doors between pairs of rooms for a number of rounds proven to be enough. Hubs don't slow it down; more rooms do, steeply.

What's measured here

• Time per layout: performance.now() around batches of generator work only (map drawing and statistics are outside the timer), divided by the layouts finished. Each method gets the same time slice per frame.
• Rejected: share of configuration-model deals thrown away because a corridor would loop back into its own room or double up an existing one.
• Fairness meter: 100% minus the distance between the histogram of "template doors kept" and the same histogram from the REFERENCE (a long-running double-edge swap chain, 100 swaps per door between samples). The striped part is the expected sampling noise. The exact check (bottom right) computes true distances on an 8-room dungeon small enough to list all 623 layouts.
• Throwing away disconnected finished layouts ("connected only") keeps the configuration model exactly fair and Curveball within its proof, over connected layouts. Rejecting moves inside a chain would not.

Where the prior technique still wins / limits

• Ordinary dungeons (?scene=normal, ?scene=big): the configuration model is exactly fair and hundreds to tens of thousands of times faster than the certified Curveball. Use it.
• One big hub (?scene=square, a 16-door square in 30 rooms): the configuration model still wins, even though it throws away more than 99.9% of its deals.
• Several hubs (?scene=hubs, raise the square slider to 18-20): almost every deal clashes and the configuration model falls behind; Curveball's cost does not change.
• Certified Curveball's cost grows with about the 4th power of the room count (80 rooms: a loading bar). The proof guarantees fairness, not speed, and only "within 1/4"; the measured distances are far smaller. The quick swap is cheapest but measurably unfair everywhere.
• Presets: ?scene=normal (20 rooms, 1-6 doors), ?scene=square (30 rooms, one 16-door square), ?scene=hubs (square + two 10-door guild halls; &square=18 for more), ?scene=big (80 rooms). Also ?frames=N, ?prewarm=MS, ?day=N, ?connected=1.