link-check accepts invalid links with unicode characters

[fredrik-bakke]

My auto-completion messed up this link and I'm surprised it passed the link-check. I guess it's a bug caused by the ℚ character but maybe it's something you want to look into.

Originally posted by @malarbol in #1424 (comment)

[VojtechStep]

For context, here's the undetected bad link from the linked PR: [sublinear rational Cauchy approximations](metric-spaces.sublinear-approximation-ℚ.md).

Okay, this was pretty tough to track down. When deciding if a link target is a file path, linkcheck runs it through http::uri::PathAndQuery::try_from. Since the library hasn't been updated for so long, it depends on http@0.2 and not the new http@1. The old one follows RFC 3986 and requires all non-ASCII characters to be percent-encoded, while the new one accepts full UTF-8 (hyperium/http#715). Since the link target is not recognized as an URL or a path, it gets categorized as "unknown". But unknown links are ignored and don't produce a diagnostic.

I'm not really sure what to do here. I don't want to add a preprocessor for this, since I don't want to parse markdown to look for links, and I don't think we have much hope of getting upstream linkcheck to update its dependencies.