It got traced to a race condition in munmap() in the kernel. The inefficient musl allocator simply triggers it more easily.
I looked at the explanation, and it seems to be at very least plausible. The kernel version bisection also lines up. We'll see if the proposed patch fixes this.