I believe it is true that this should never happen. To reproduce, create a pyproject.toml with:

[project]
name = "uv-lock-instability"
version = "0.1.0"
description = "whatever"
requires-python = ">=3.9.0"
dependencies = [
    "jax==0.4.17",
]

And now do a lock followed by another lock:

$ rm -f *.lock
$ cargo r --manifest-path ~/astral/uv/Cargo.toml -p uv -- lock
$ cp uv.lock before.lock
$ cargo r --manifest-path ~/astral/uv/Cargo.toml -p uv -- lock
$ cp uv.lock after.lock
$ diff before.lock after.lock
15c15
<     { name = "zipp" },
---
>     { name = "zipp", marker = "python_version < '3.10'" },

The diff output at the end should be empty, but it isn't.

Running the same steps in the ecosystem/transformers directory produces a larger diff (and this MRE was minimized from that).