object.extend(mod) is occasionally nondeterministically applied

[halostatue]

NoMethodError: undefined method `load_preferred_extension' for an instance of MIME::Types
    lib/mime/type/columnar.rb:55:in `preferred_extension'
    lib/mime/type.rb:221:in `__extension_priority_compare'
    lib/mime/types.rb:156:in `block in type_for'
    <internal:core> core/array.rb:1588:in `isort_block!'
    <internal:core> core/array.rb:1498:in `mergesort_block!'
    <internal:core> core/array.rb:1406:in `sort_fallback'
    <internal:core> core/array.rb:1707:in `sort'
    <internal:core> core/array.rb:1707:in `sort!'
    <internal:core> core/enumerable.rb:470:in `sort'
    lib/mime/types.rb:151:in `type_for'
    lib/mime/types/registry.rb:36:in `type_for'
    test/test_mime_types.rb:16:in `block in mime_types'
    <internal:core> core/kernel.rb:520:in `tap'
    test/test_mime_types.rb:8:in `mime_types'
    test/test_mime_types.rb:37:in `test_0004_is countable with an enumerator'

The line numbers against mime-types/ruby-mime-types main differ from the exact report here (as I have been investigating this for a couple of days), but are regularly but inconsistently reproducible with rake test using truffleruby 24.2.1, like ruby 3.3.7, Oracle GraalVM Native [arm64-darwin20]. Sticking with a given seed (SEED=6845 rake test) does not change the behaviour to deterministic, but roughly ⅓ to ½ of the test runs fail.

I have tried several different resolutions: removing metaprogramming, putting the container.extend(MIME::Types::Columnar).load_base_data(path) call inside of a Mutex.new.synchronize { … } block, etc. None have worked. There is one more change that I can make, but I wish to avoid that, because it will take a couple of weeks to resolve for a single interpreter misbehaving.

When a test is run by itself, this rarely happens. But it happens fairly often when the test suite as a whole is run.

[fniephaus]

Thanks for the report, @halostatue.

@nirvdrum any idea what is going on here?