The shift roster builder checks its rosters with a chain of lower bounds. The roster is proven when one of them reaches its score. One of those bounds comes from column generation, and a test makes sure that column generation cut off early never claims a bound above the best roster.
That test took 61 seconds, the slowest single test in the suite. The phase timer in the solver's output (timings) showed why. Each "cut short" call spent 9.3 of its 10.6 seconds in the exact search, which runs after the column generation under test. With the proof left open, the search ran to its default 8-second limit on a week the test never looked at again.
The fix was one option on the test's cut calls, exactMs: 300. Shrinking a budget can hide a bug, so before changing it we checked three things:
- Both cut-short passes still run. One runs inside the exact search (an "open-hours row" pass). At 300 ms it still ran on all six calls, as it did at 8 seconds. At 1,000 ms it did not, because some first passes then finished on their own. So the shorter budget is not the larger one here.
- Planted bugs are still caught. Raising either pass's bound by a billion makes the new test fail. The test now also checks the inner pass's bound directly, so that failure names the pass.
- The exact search keeps its own tests, elsewhere in the suite.
The test now takes 13 seconds and the file 30 instead of 76. The solver itself did not change. The rest of the suite's slow tests are mostly testing exactly the budgets they spend, so the next win will not be this easy. The full reading is in docs/history/ci-perf-2026-11-16-shift-roster-cg-hardstop-exact-ms.md.