fix: handle General/Complex Versioning in --bump

[liskin]

fix: handle General/Complex Versioning in --bump

The previous fix to handle non-semver versions in --bump used a crude heuristic - a check whether there's a "." in the version string - to decide whether to attempt to chunkify or just fall back to returning the new version unmodified. This, however, resulted in no bump happening (None returned from check_semver_bump) when either of the old/new versions looked like "v1.2.3".

Additionally, versions like "1.2a1" and "1.2a2" were considered the same, because: a) the heuristic failed to recognise they won't be parsed/chunkified correctly, and b) chunkify used nth that silently throws parts of the version away.

I'm proposing to fix this by reimplementing chunkify using the much more generic versions::Mess format, which simply splits the version number into chunks and separators. We then convert it into a simpler format of just chunks (first chunk as is, further chunks with a separator prepended).

This design has an issue: we don't recognise a change in the versioning scheme. A bump from "<commit sha>" to "v1.2.3" will result in just "v1", because the commit sha is parsed as a single chunk. The most obvious case of this, "latest" being parsed as a single chunk, is handled explicitly in the code, as it would just be a mess otherwise.

A potential workaround for this issue would be to add a flag (e.g. --pin) that would make --bump skip the check_semver_bump logic and always use the full new version (as suggested in https://github.com/jdx/mise/issues/2704#issuecomment-2399624421). This would also help in the following case: a project using variable length version numbers instead of the full 3-chunk semver. Trying to follow this sequence of bumps: "20.0", "20.0.1", "20.1" isn't possible with the current logic.

Related: 0b2c2aa ("fix: upgrade --bump with non-semver versions (#2809)")

fix: Allow --bump from 20.0.1 to 20.1

It's weird so we still warn, but returning None from check_semver_bump only makes sense if the versions are deemed to be the same. Otherwise it's just confusion for the user — the UI presents this as an upgrade, proceeds to uninstall the old version, but fails to do the actual bump and no new version is installed.

[liskin]

This design has an issue: we don't recognise a change in the versioning scheme. A bump from "<commit sha>" to "v1.2.3" will result in just "v1", because the commit sha is parsed as a single chunk. The most obvious case of this, "latest" being parsed as a single chunk, is handled explicitly in the code, as it would just be a mess otherwise.

The tests caught this, precisely because it now tries to bump dummy from "ref:master" to "2.0" (the new/latest version being "2.0.0").

I suspect that a proper solution would require us to detect whether the two versions follow the same/similar versioning scheme, and don't try to do check_semver_bump if old is something like "ref:4a2d75f…" or "prefix:20". It's not trivial, though, as there's some inconsistency here – the pipx/uv backend doesn't support ref: or rev: at all, for example, it expects to just get the commit sha directly. And also, list_outdated_versions tries to some prefix stripping magic—although maybe that's not necessary now that we handle non-semver versions?

Any ideas?

[jdx]

I think right now it probably doesn't work right with ref: or prefix: so we should probably just ignore those in mise up altogether, maybe with a warning, until we have a better solution for them.

[jdx]

are you actually seeing v prefixed versions somewhere? we should be removing that prefix everywhere we can by this point

[liskin]

They're in my config so I thought it'd be nice if mise upgrade --bump didn't choke on them. I can remove them manually but it'd be nice if others didn't need to.

[liskin]

But you've got a point that the test cases don't reflect reality. There might be a "v" in old but not in new.

[liskin]

I think right now it probably doesn't work right with ref: or prefix: so we should probably just ignore those in mise up altogether, maybe with a warning, until we have a better solution for them.

Actually, the code already kinda tries to ignore them:

mise/src/toolset/mod.rs

Lines 402 to 420 in 9bedf17

	match out.tool_request.clone() {
	ToolRequest::Version {
	backend,
	version: _version,
	options,
	source,
	} => {
	out.tool_request = ToolRequest::Version {
	backend,
	options,
	source,
	version: out.bump.clone().unwrap(),
	};
	}
	_ => {
	warn!("upgrading non-version tool requests");
	out.bump = None;
	}
	}

But it's half-assed — only skips the bump, but still performs the uninstall. And then, as I mentioned earlier in https://github.com/jdx/mise/issues/2704#issuecomment-2399555891, the dry-run output doesn't match what's actually going to happen, so the output in the failing test is "Would install dummy@2.0.0" but that's not what would happen in practice, as there'd be no bump.

Oh, what a can of worms. I think I give up for today :-D

[jdx]

I did some work around prefix: in #3054