Finally figured it out:

1. musl is reclaiming space from the executable starting at offset 0x224B20, i.e. at the end of the bss

2. this reclaimed space gets used for the dso struct of the first shared lib

3. the last variable in the bss appears to be scratch space for checksum computation

4. the code is assuming "unsigned long" to be 4 bytes, which isn't the case on 64bit platforms

5. the checksum code overflows out of the bss, corrupting the dso struct

6. this issue is masked in a glibc environment because the loader doesn't make the unused part of the program pages available to malloc.

7. valgrind doesn't catch the problem because it doesn't bound-check globals

Sorry about the noise.