Or a C implementation at https://tromp.github.io/pearls.html#sieve which runs in well under 10s.

I'd be interested in seeing an explanation of the code, since it looks pretty incomprehensible to me. Per the arbitrary rules I set for myself, I'm not allowed to precompute/hardcode the wheel (looks like this implementation uses a hardcoded wheel of size 2x3x5=30). I wonder if/by how much the performance would suffer by computing and storing the coprime remainders in memory instead of handing them directly to the compiler.