1.95 beta regression: "malformed `feature` attribute input"

[theemathas]

Regression found in the 1.95.0 beta crater run.

[INFO] [stdout] error[E0565]: malformed `feature` attribute input
[INFO] [stdout]   --> /opt/rustwide/cargo-home/registry/src/index.crates.io-1949cf8c6b5b557f/assemble-core-0.2.0/src/unstable.rs:16:5
[INFO] [stdout]    |
[INFO] [stdout] 16 |     #[feature(feature = "text_factory")]
[INFO] [stdout]    |     ^^^^^^^^^^^^^^^^^^----------------^^
[INFO] [stdout]    |     |                 |
[INFO] [stdout]    |     |                 didn't expect any arguments here
[INFO] [stdout]    |     help: must be of the form: `#[feature(feature1, feature2, ...)]`

• https://crater-reports.s3.amazonaws.com/beta-1.95-1/beta-2026-03-05/reg/assemble-std-0.2.0/log.txt

[INFO] [stdout] error[E0539]: malformed `feature` attribute input
[INFO] [stdout]  --> src/models/whisper/mod.rs:1:1
[INFO] [stdout]   |
[INFO] [stdout] 1 | #![feature("whisper")]
[INFO] [stdout]   | ^^^^^^^^^^^---------^^
[INFO] [stdout]   | |          |
[INFO] [stdout]   | |          expected a valid identifier here
[INFO] [stdout]   | help: must be of the form: `#![feature(feature1, feature2, ...)]`

• https://crater-reports.s3.amazonaws.com/beta-1.95-1/beta-2026-03-05/gh/MikeIvanichev.norma/log.txt
• https://crater-reports.s3.amazonaws.com/beta-1.95-3/1.95.0-beta.3/reg/norma-0.0.3/log.txt

[INFO] [stdout] error[E0539]: malformed `feature` attribute input
[INFO] [stdout]  --> src/lib.rs:4:1
[INFO] [stdout]   |
[INFO] [stdout] 4 | #[feature("serde_feature")]
[INFO] [stdout]   | ^^^^^^^^^^---------------^^
[INFO] [stdout]   | |         |
[INFO] [stdout]   | |         expected a valid identifier here
[INFO] [stdout]   | help: must be of the form: `#[feature(feature1, feature2, ...)]`

• https://crater-reports.s3.amazonaws.com/beta-1.95-1/beta-2026-03-05/reg/isfet-0.1.3/log.txt

[INFO] [stdout] error[E0539]: malformed `feature` attribute input
[INFO] [stdout]  --> src/lib.rs:4:1
[INFO] [stdout]   |
[INFO] [stdout] 4 | #[feature("proc_macro_lib2")]
[INFO] [stdout]   | ^^^^^^^^^^-----------------^^
[INFO] [stdout]   | |         |
[INFO] [stdout]   | |         expected a valid identifier here
[INFO] [stdout]   | help: must be of the form: `#[feature(feature1, feature2, ...)]`

• https://crater-reports.s3.amazonaws.com/beta-1.95-3/1.95.0-beta.3/reg/starlane-primitive-macros-0.3.21/log.txt

In four out of the five affected crates, it seems like the authors of the code intended to write cfg(feature = "something"). In starlane-primitive-macros, it seems the author wrote #[feature(....)] when they meant #![feature(....)], with a !

[theemathas]

Presumably regressed in #152985. cc @JonathanBrouwer @jdonszelmann

[apiraino]

Assigning priority (discussion on Zulip).

@rustbot label -I-prioritize +P-medium

[JonathanBrouwer]

The regression is that the arguments of feature attributes on invalid targets used to be unchecked... ugh
Minimum reproducer:

#[feature("meow", this, was, not, checked)]
mod x {}

We have the following options

1. Just leave this unchanged, accepting the breakage
2. Change this back, then revert with T-lang approval
3. Change this back

Given that this is effectively a bug fix, the impact is relatively small, and unfixing this bug is a little annoying, I'd prefer option 1, but if anyone does not like this we can go for option 2 instead

cc @rust-lang/lang

[jieyouxu]

The calculus is basically:

• This should've been rejected but wasn't because it was too lax (intended to be malformed). The attribute PR just starts actually doing the checking.
• We don't see evidence of widespread ecosystem breakage, the affected crates AFAICT are not widely used (https://github.com/assemble/assemble-core, https://github.com/MikeIvanichev/norma, https://crates.io/crates/isfet)
  • assemble-core last updated 10 years ago
  • norma probably just a tiny diff (MikeIvanichev/norma@aa8289c) should be feature = "whisper" instead
  • isfet looks toy-ish
  • starlane-primitive-macros 0.3.21 is https://crates.io/crates/starlane-primitive-macros/0.3.21, repo https://github.com/starlane-io/starlane, 1.3k cumulative downloads, also looks like a toy repo

I would recommend

1. Just leave this unchanged, accepting the breakage

with an explicit T-lang FCP. Or just let it slide with a compat relnotes if T-lang thinks that's fine.