This is intentional. We want to wait for both tasks to finish and let result = result only check for the one that actually returns result. We also need to wait for the writer task which doesn't return anything.

Oh, right. In that case, how about using async tasks instead of a task group? That would give you a cleaner way to wait for both with no force unwraps. Both things continue to run in parallel and guaranteed to wait for everything to complete.

async let buffer = {
    var buffer = Data()
    for try await chunk in standardOutput {
    let currentChunk = chunk.withUnsafeBytes { Data($0) }
        buffer += currentChunk
    }
    return buffer
}()

_ = try await standardInputWriter.write(Array(expected))
try await standardInputWriter.finish()

return try await buffer

Ditto. This is intentionally to wait for both tasks.

Should registerFileDescriptor precondition that the fd is not already in the _registration map?

We can, but I don't think it's necessary because epoll_ctl will fail if it's already registered so you get the same error.

✘ Test stressTestWithLittleOutput() recorded an issue at IntegrationTests.swift:1460:6: Caught error: An error occurred within the AsyncIO subsystem: failed to add 27 to epoll list. Underlying error: UnderlyingError(rawValue: 17)

Could this be because _safelyClose does not call AsyncIO.shared.removeRegistration for the fd before closing it? Reading I/O fails because of low limit, fails, calls _safelyClose, fd is closed, fd # gets reused, tries to re-add, already in the epoll list, boom.