There may be "gigabytes of external libraries" included in the OS, but these are necessary for using current-gen hardware. Libraries and assets that are not just for interfacing with the hardware and basic stuff like opening a window are rarely taken advantage of, except sometimes for text rendering. System decompressors used to be used, but nowadays, productions come with their own specialized unpackers (ex: kkrunchy) that are much more powerful yet smaller than anything that comes with the OS. EDIT: I see MIDI being mentioned, and it is indeed still used. But for 64k, there are softsynths that sound way better than system MIDI and only take a small part of the size budget, sometimes going as far as physically modelling instruments. 64k is crazy.
These 64k intros could be made run on the bare metal with not much difference in size if there was hardware for it. But the way modern hardware is designed, it is impossible. If you really insist on going "bare metal", try an Amiga ;)
I found this: https://github.com/nanochess/RayTracer
Basically, you can use the GPU and you can use most of the 64kB to generate procedural content.
Everyone gets the same base OS and compete against _each other_, so nobody gets a special advantage. If someone finds a clever way to use the base OS for more bytes (such as reusing the standard Windows fonts or MIDI soundfont or CAB compressor), then so does everyone else, and they use the saved space for even more cool and impressive stuff next release.
Here's some tricks used for 4kb demos (some of which are applicable to 64kb demos too):
https://in4k.github.io/wiki/aulds-4kb-introduction
https://in4k.github.io/wiki/aulds-4k-synth
https://in4k.github.io/wiki/exe-packers-tweakers-and-linkers