Do not report a channel closed underneath us as EBADF on close - #9723

Fixes #9722.

Thread#kill unblocks IO by interrupting the native thread, and the JDK closes the interruptible channel in response. Ruby never closed that stream, but the ensure that follows calls IO#close, and ChannelFD.finish() raised ClosedChannelException — surfacing as Errno::EBADF, which Thread#join then re-raises into the caller.

The close is now skipped when the channel has already gone. Two details:

• isOpen() guards ch.close() only, not the branch, so the fileno bookkeeping still runs. It cannot be dropped either: FileDescriptorByteChannel.close() throws on a second close, and IOChannel.close() would call a user object twice.
• unregisterWrapper now takes the wrapper and removes only if the fileno still maps to it. By key alone, a stale close evicts a live wrapper on a recycled descriptor and a later close closes an unrelated file.

The one-argument unregisterWrapper(int) is left in place, now unused and unsafe unless the caller owns the mapping — deprecating it seemed an API call to leave to you.

Evidence

Forcing test red before, green after. 800 reproduction rounds: 14 failures on Java 21 and 12 on Java 25 before, 0 after; CRuby 0 in 400. MRI net/http reproduces the defect on base at about 1 run in 10, absent across 10 runs after. test_io.rb 49, MRI ruby/test_io.rb 200, stdlib net socket io fiber 362 — all green; io, file and socket specs identical.

Limits

Only close is fixed. After the interrupt the stream still reports closed? == false and still fails write and select. And io.to_channel.close; io.close no longer raises Errno::EBADF, which CRuby does — a lost diagnostic.

The cause is the interrupt itself, as the FIXME at OpenFile.java:1466 says. Giving writeInternal the pre-select readInternal has would fix all four, at the cost of a select per write. Happy to take that on if you want it.