The free shift roster builder accepts up to 60 people and 400 shifts. We had only tested hostile files that were refused; we had not timed a week that is allowed but large. We built one: 60 people, 56 shift lines a week, seven people needed on each, 392 shifts in all.
It ran for more than 110 seconds before we stopped it, printing nothing. The solver is a chain of passes (annealing, ejection chains, twelve plateau walks, an exact check), and each is tuned to keep going while any shift is open. On a week where people are the limit, some shifts will always be open, so every pass ran in full.
The fix is a wall-clock budget. Weeks with more than 150 shifts get 20 seconds of improvement passes; after that the best roster so far is returned and the result says budgetHit: true. Smaller weeks, including every guide page example, run as before. The same 392-shift week now finishes in 21.7 seconds with 300 shifts covered, which is all 60 people can work at five shifts each.
What we have not shown: that 20 seconds loses nothing on big weeks that are coverable. On the one week we timed, capacity was the limit, so the budget cost nothing. A slower machine also stops earlier. If you run the builder on your own server, timeBudgetMs sets it, and 0 turns it off.