Documented `skip_children` workaround for `mod` walking seems to be ignored on stable rustfmt

[martingms]

I think after @UebelAndre's #3983 (which closed #3850), the rustfmt_srcs_aspect now picks up rust_library targets whose srcs mix hand-written and generated files, where it didn't before. In my repo this surfaces a failure on a common shape: a hand-written lib.rs declaring mod generated; where generated.rs comes from a genrule. The aspect runs rustfmt and fails with:

This seems to be the case docs/rust_fmt.vm already calls out:

Any target which uses Bazel generated sources will cause @rules_rust//tools/rustfmt to fail with failed to resolve mod MOD. To avoid failures, skip_children = true is recommended to be set in the workspace's rustfmt.toml.

The thing is, that workaround doesn't seem to actually do anything on stable rustfmt anymore. With skip_children = true in rustfmt.toml, stable rustfmt prints:

Warning: can't set `skip_children = true`, unstable features are only available in nightly channel.

and then walks the mod declarations anyway, producing the failure above. So as far as I can tell the documented mitigation is a no-op on stable.

Curiously, passing the same option as a CLI flag seems to work though:

rustfmt --config skip_children=true lib.rs   # exits 0, mod walk skipped

Is there some other known workaround?

[martingms]

The following patch seems to work for us, for what it's worth.

--- a/rust/private/rustfmt.bzl
+++ b/rust/private/rustfmt.bzl
@@ -87,6 +87,7 @@
     args.add(rustfmt_toolchain.rustfmt)
     args.add("--config-path", config)
     args.add("--edition", edition)
+    args.add("--config", "skip_children=true")
     args.add("--check")
     args.add_all(srcs)