artifact-dependencies cause entire workspace to be nightly-only

[tamird]

Problem

Suppose I have a workspace:

.
├── Cargo.toml
├── my-crate/...
├── support-crate/... # contains publish = "false"
│   └── src/
│       └── main.rs
├── integration-test/
    └── Cargo.toml # contains publish = "false" *and* build-dependencies.support-crate.artifact = "bin"

Before the use of "artifact" I could do various things with stable cargo, even if my integration tests are nightly-only. After introducing the use of "artifact", the entire workspace is unusable with stable cargo:

$ cargo +stable check
error: failed to load manifest for workspace member `.../integration-test`

Caused by:
  failed to load manifest for dependency `support-crate`

Caused by:
  failed to parse manifest at `../support-crate/Cargo.toml`

Caused by:
  `artifact = …` requires `-Z bindeps` (bpf-linker)

We should allow folks to partially opt in, else hardly anyone will be willing to put mileage into this feature.

Proposed Solution

No response

Notes

No response

[epage]

This was discussed in Zulip: https://rust-lang.zulipchat.com/#narrow/stream/246057-t-cargo/topic/unstable.20cargo.20in.20workspace.20breaks.20stable.20workflows

Quoting myself

I believe the primary challenge is to that to fully resolve the dependency tree, all workspace members have to be loaded.
This case particularly highlights the problem if we just said "ignore the nightly features and parse to get dependencies" because the dependency is a nightly feature.
Unsure what the complexity or ramifications are to skipping resolving the package.

• Editing the package should cause changes for the rest of the tree even if the package wasn't selected to be built
• If a different package changes, then we need all packages to evaluate

Maybe if we checksumed packages (taking into account workspace inheritance) and said that if anything needs resolving, we will then try to load all workspace members to resolve everything