apple_incremental_game: 8 cases at once on 4 CPUs move deadline solvers by 3-11 points #44

Task

apple_incremental_game

Claim

The judge runs 8 cases at once on the official 4 CPUs (max_workers=min(8, len(case_files))), so each case gets about half the CPU that 4 at a time would give it. SCHEMA says a parallel grader "sizes the pool with SFORGE_CPUS, not nproc". The pool mechanism is the one in #40.

Impact

A solution that reads the clock searches less: seven real files gain 3.65-10.86 of 100 at 4 at a time (14 pairs, 4 vCPUs online, one judge), and two with their own cutoffs at 22 s and 25 s don't change. None loses a case on a quiet host, and all 89 grades in our agent runs passed 50 of 50. A solution with a fixed cost above about 15 s of CPU per case times out: our control gets 2 of 50 OK and score 0. We hold no real solver that does this at official settings; the Regression summary's "historical 1 (no time check)" and Known issue 3's submission may.

Evidence and how to reproduce

The control and the commands are in #41. Community at f92857a gives 2 of 50 OK, score 0; 4 at a time gives 50 of 50 OK, score 5.04 (n=2 per judge).

Proposed fix

min(8, SFORGE_CPUS, cases), in #41.

Limits

Part of Known issue 3, which stays: limits are wall clock. Whether this is a scoring defect or a resource-policy change depends on whether the archived submissions above lose cases on your host. We can't test that.