Add a new unstable flag -Ztreat-pub-as-pub-crate

[mu001999]

This PR adds a new unstable flag -Ztreat-pub-as-pub-crate as @Kobzol suggested.
When compiling binary crates with this flag, the seed worklist will only contain the entry fn and won't contain other reachable items. Then we can do the dead code analysis for pub items just like they are pub(crate).

Related zulip thread #general > pub/pub(crate) within a binary is a footgun.

[rustbot]

r? @jdonszelmann

rustbot has assigned @jdonszelmann.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[Noratrieb]

What are the plans for this flag? Being made the default? As just a flag, it's essentially useless, because no one will know about it. So I'd be opposed to just adding it without any plan for making it useful for everyone.

[mu001999]

Maybe this could be a separate lint, I haven't thought too clearly yet

[jdonszelmann]

Yea I do agree with nora here, maybe open a (specific) zulip thread to make a proper plan for this?

[jdonszelmann]

I guess I'd prefer this as a lint (attribute at the crate root) myself

[jdonszelmann]

@rustbot author

[rustbot]

Reminder, once the PR becomes ready for a review, use @rustbot ready.

[Kobzol]

I guess that the use-case that I'm envisioning is to have an opt-in way to treat pub stuff as pub(crate) in binary crates. If it was an allow-by-default lint, which could be manually denied, that would serve that use-case, I think.

[GrantGryczan]

Why would the end goal be for this to be opt-in? I and others want to see #74970 become a default. I wouldn't opt into this, because that would make my codebase less idiomatic, diverging from how everyone else who hasn't opted in uses pub.

The reason this is useful as a default is because it doesn't semantically make sense to distinguish between pub and pub(crate) in binary crates. This creates a footgun because, for example, pub items can't receive dead_code warnings. If it's just opt-in, then that doesn't remove the footgun for anyone who hasn't opted in, and I don't see a benefit to pub and pub(crate) being distinct in binary crates.

The only reason I can see that someone might not want to opt into this is that one might use pub to prevent dead_code warnings as opposed to pub(crate). But I think this intent might be better communicated using #[expect(dead_code)] rather than the implicit behavior of pub. Dead code detection is useful as a default. (The only downside to that is that you then have to remove the #[expect(dead_code)] if you later use it, but I'll leave it to others to decide if that coupling is an important consideration.)