$ muon samu -j 0
Segmentation fault

ninja supports -j 0 to mean "start unlimited jobs in parallel", analogous to make -j. samurai sets buildopts.maxjobs = -1 in this case (e.g. SIZE_MAX).

However, since muon's samurai allocates all maxjobs upfront, this causes an attempted allocation of SIZE_MAX jobs. Even though samu_arena_alloc can't fail, SIZE_MAX * sizeof(jobs[0]) is larger than SIZE_MAX, so samu_reallocarray fails and returns NULL. Then, when samu_build() tries to access the NULL array, it crashes.

Here are some possible solutions:

1. Don't support -j 0. I'm not convinced that unlimited parallelism is a good feature, so it's probably fine to not support it (change num < 0 to num <= 0 in samu_jobsflag()).
2. Since muon samurai uses it's own job spawning code, it doesn't have to keep a pollfd array, so you could allocate the job structs separately and make jobs a linked list.

In either case, I think the fatal should be added back to samu_xreallocarray, since callers expect that it always succeeds.