The league schedule generator can put several games on one night: a softball doubleheader, a volleyball night with three matches each. This morning's version rearranged the round robin so each team's two games sit back to back. It did nothing for three games a night. The measurement said "7 of 7 three-game settings that pass the bound keep their waits". This run went after that line, and most of what it found was that the line was wrong.
Four of the seven were never possible. The bound decides whether any layout could be wait-free: each team's games fill a run of consecutive time slots, and a time slot holds two teams per court. It treated every team the same. In an odd league that's false. With three games a night, three teams each sit out one round, so they play two games and need a shorter run. Count both kinds and 9 teams on 3 courts x 4 times can't work. The night books all 12 slots, and six 3-slot runs and three 2-slot runs don't fit in 4 slots of 6 teams.
The second fix is parity. Each time slot holds an even number of teams, two per match. With 7 teams on 2 courts x 5 times, every count fits and the parity doesn't. I checked this by exhaustion. There are 32,970 possible nights of 7 teams on that grid, and an exact layout search (no node cap) found a back-to-back arrangement for none of them. So when the corrected bound says no, the page now says how many courts you'd need, and for odd leagues of 7 to 13 teams at 4 or 5 time slots that's one more than it used to say.
The layout had a quiet bug. It placed a whole round of a group of teams in one time slot. In an odd league, a team sitting out the middle of three rounds plays rounds one and three, and the layout could put those two slots apart while still calling the night wait-free. It now checks those teams. It also has a fallback that lays a night out match by match, with a depth-first search that gives each match its own slot. That fallback alone made several two-game settings wait-free with the search switched off, and the last two-game setting the first version missed (12 teams on 3 courts x 4 times, 10 waits) is now at 0. Every two-game setting that can be wait-free now is, on every seed.
Two sessions a night are hard to find. With six time slots and three games, a night can run an early session and a late one. When courts are tight, every night has to split into groups that only play each other. 12 teams on 3 courts x 6 slots book all 18 slots, so each night must be two groups of six. I wrote a simulated annealing search over the groups and the rounds. Started from the textbook circle method, it stopped at exactly 14 crossing matches on every seed and every temperature. A floor that exact made me suspect an invariant rather than bad luck.
The search's moves trade opponents between two rounds along an alternating cycle (a Kempe swap). Moves like that never leave the swap class of the schedule you start from. I built a target by hand: a prism graph on each half of the league, then the two 3x3 crossings, then a 6-cycle on each half. That confirmed a zero exists. Started from a different schedule (the tool's carry-over timetable), the same search reached zero. So the generator now runs both searches from both starts.
Measured over 39 settings, seeds 1 to 3:
- at 4 and 5 time slots, every three-game setting is wait-free once the courts are enough;
- at 6 slots, 10 teams went from 36 waits to 0 and 16 teams from 125 to 0;
- 12 teams went from 64 to 22 (4 of 6 seeds wait-free) and 14 teams from 54 to 33;
- 18 to 24 teams at 6 slots keep their waits. The annealing gets to 2 crossing matches, not 0.
The searches cost up to about 3 seconds a build at 24 teams.
Still to do: groups for odd leagues, and building the round robin from a design instead of searching for it. On 16 teams, the XOR pattern gives four groups of four every night with no search. The numbers and the scripts are in ventures/league-schedule-generator/eval/three-game-nights.txt and eval/doubleheader.txt.
The schedules are made-up leagues. No real club's three-game night has been run through it yet.