Looking back at the original ticket issue, I'm not yet convinced that delaying the errors is the right approach. What if we allow users to just list the directories they want to be included? I.e.

(include_subdirs
 (kind qualified)
 (dirs * \ bar-baz))

I think it would be a smaller change, and also a useful other in other contexts.

I'm not yet convinced that delaying the errors is the right approach.

I disagree because of the case in #7628:

$ tree
.
├── foo
│   └── bar.ml
├── invalid-module-name
│   └── opaque.txt
├── dune
├── dune-project
└── x.ml

IMO there's no reason why the above should fail if you have an unrelated directory that won't contain OCaml sources.

I think specifying directories renaming module interfaces is definitely useful, but I consider it to be an additional feature, not a replacement.

So if you were to modify #7628 so that invalid-module-name contained an .ml file, the error would not be delayed? By delayed, I mean that building some other target in the same directory as the library would not trigger this error.

So if you were to modify #7628 so that invalid-module-name contained an .ml file, the error would not be delayed? By delayed, I mean that building some other target in the same directory as the library would not trigger this error.

I pushed d8dd664 to show that the error isn't delayed in that case.

To make that test-case more bullet proof, could you build something trivial defined with a rule stanza instead of an empty library? It probably won't make a difference, but I'd be a bit more confident.

Just pushed a commit doing that