Adam Dinwoodie, preparing Nixpkgs for Git v3, found that a Git built with WITH_BREAKING_CHANGES=1 still creates SHA-1 repositories with git init. He confirmed this at v2.56.0, on next and on main, and asked whether that was expected.
Junio Hamano looked at setup.c:repository_format_configure(). He said REPOSITORY_FORMAT_INIT sets the hash algorithm to GIT_HASH_DEFAULT, but read_repository_format() forces it to SHA-1 when the version is -1 (no config yet), so new repositories kept getting SHA-1. He also described a latent bug: the hash-algo block guards only the GIT_DEFAULT_HASH branch with a version check, leaving the branch that uses the configured hash unguarded, unlike the ref-storage block below it.
In response to a patch from brian m. carlson, Junio asked whether it breaks re-running git init in an already initialised repository. He wrote a test modelled on the patch's new one: create a repository with the non-default hash, then run git init with no hash specified. He called it debatable what Git 3.0 should do there, and listed possibilities including converting an existing SHA-1 repository to SHA-256 or leaving it alone.
carlson said a "limbo" state, where a repository's hash is undecided until needed, was discussed. It would require protocol changes that nobody has implemented; he thought Peff had considered it but no decision was reached. He said a repository would still need an initial hash for local work, including bare repositories, so limbo would be a separate option, perhaps extensions.potentialObjectFormat, and the patch should go in either way.
The re-init behaviour Junio raised does not appear to be settled in the excerpts.