Interception of localhost traffic is in fact a non-issue, someone able to do it can do worse. So TLS really shouldn’t be necessary on localhost, that’s it.
Certainly seems likely to be the case, but I'm open to there being a theoretical attack that involves being able to inspect the data flowing through the TCP stack without being able to inspect the process spaces of the two endpoints. E.g. if the TCP stack was able to be put in a debug mode that was logging packets somewhere, you'd prefer those to be encrypted. It's pretty far-fetched in terms of an attack surface, but "this data never exists unencrypted outside these two processes' memory spaces" is objectively stronger in a specific way than plaintext transiting kernel buffers.
