Fix repo contents cache FileValue staleness

[fmeum]

Ensures that files under repo contents cache entries are not reported as missing after the cache has been deleted while the Bazel server is running. See the long comment in RepositoryFetchFunction for why this happens and how it is fixed.

Fixes #26450

[Wyverald]

I still don't fully understand which part of this PR made the repo contents cache dir special. Is it just the DirtinessCheckerUtils::isCacheableType change?

[fmeum]

re the second point: the original comment says "The only way other than external repositories for external source files ...". That doesn't seem wrong? The repo contents cache is only usable through external repos after all.

That depends on how strictly one interprets this: I read it as "The only way other than external repository directories...", which used to be true, but no longer is since the repo contents cache entries don't live under such directories (although of course they are reachable through them via symlinks). Ignoring that part, the conclusion "So the set of external files is small." is definitely wrong now since the repo contents cache file contributes a large number of EXTERNAL_OTHER files.

[fmeum]

I still don't fully understand which part of this PR made the repo contents cache dir special. Is it just the DirtinessCheckerUtils::isCacheableType change?

It's really just that (plus the switch to FileStateValue).

[Wyverald]

re the second point: the original comment says "The only way other than external repositories for external source files ...". That doesn't seem wrong? The repo contents cache is only usable through external repos after all.

That depends on how strictly one interprets this: I read it as "The only way other than external repository directories...", which used to be true, but no longer is since the repo contents cache entries don't live under such directories (although of course they are reachable through them via symlinks).

I think that's a bit uncharitable -- I don't think we call $outputBase/external "external repository directories". "External repositories" is simply the name of the feature that allows you to define repos etc. But that doesn't really matter because...

Ignoring that part, the conclusion "So the set of external files is small." is definitely wrong now since the repo contents cache file contributes a large number of EXTERNAL_OTHER files.

That I agree with. In fact it might have already been wrong before; any files pointed to by a local_repository would have been EXTERNAL_OTHER, and this would have included bazel_tools.