Fix edge cases in lockfile handling

[fmeum]

• Don't run the core logic when --lockfile_mode is off or error but the command doesn't forward options to Skyframe.
• Honor reproducible per extension eval factor, not per extension.
• Fix encoding conflict between isolation key and use_repo_rule's fake extension names

This doesn't require a lockfile version bump as use_repo_rule's fake extension (so far) isn't included in the lockfile.

Work towards #24723

[fmeum]

@bazel-io fork 8.0.1

[fmeum]

@bazel-io fork 7.5.0

[fmeum]

@meteorcloudy Tests are passing now.

[meteorcloudy]

Why do we need this change?

[fmeum]

This is for the third point in the description: module extension IDs with an isolation key and those for repo rules would conflict after serialization into a string (for the lockfile), which results in a crash when parsing them back.

[Wyverald]

this seems dangerous to me. Changing the separator to + here would mean that (IIUC) we can't have it in the .bzl file label, so essentially using any repo rule from a non-main repo would fail (because that would result in an extension name like @@foo++foo_ext+foo_repo//:bar.bzl+my_repo_rule). In fact I'm surprised all tests passed -- maybe we never use_repo_rule from non-main repos in our tests?

(and yes, I understand that this also means % was never allowed to show up in the .bzl file label, but this was also true of aspects etc.)

[fmeum]

Hmm, why exactly would this fail? The synthetic extension name as passed in by module file evaluation is "prettified" to just _repo_rules for the purpose of generating the extension unique prefix: https://cs.opensource.google/bazel/bazel/+/master:src/main/java/com/google/devtools/build/lib/bazel/bzlmod/BazelDepGraphFunction.java;l=186?q=_repo_rules&ss=bazel%2Fbazel&start=41.

[Wyverald]

Yes, the repos generated by an innate extension have the form of module+_repo_rulesN+repo_name. But the name of the extension corresponding to the repo name prefix module+_repo_rulesN is currently <bzl_file_label>%<rule_name> (code). This PR is changing that line to split on +, which should result in errors.

[meteorcloudy]

Thanks!