The CI is fixed. The issue was actually with enc_arg in transcode.c:

Encoding::Converter.asciicompat_encoding
    econv_s_asciicompat_encoding
        enc_arg
            rb_to_encoding_index
                rb_enc_find_index
                    load_encoding // creates thread when multi ractor     

So, that test was creating 7x500=3500 threads, and was waiting for them sequentially. I lowered it to 7x50=350 threads, but even that might be high.

I tried removing the call to load_encoding in rb_enc_find_index, because I don't like it that we try to load encodings that aren't registered. It's been there since 2008 (0052259), and when I removed it test-all still passed. It's possible that all the other test suites also pass without it. We should at least have a test case that covers this behavior.

We could look into removing it, but I'm not sure that's possible as it would be a breaking change.