kernel: implement FIFO (named pipe) semantics — fixes process substitution / brew config

[brandonpayton]

Why

A FIFO (first in, first out), also called a named pipe, lets one process read a byte stream written by another process through a filesystem path. EOF means end of file; a reader must not see EOF while a writer can still connect and produce data.

mknod(S_IFIFO) / mkfifo created a regular file with no pipe semantics — the mknod dispatcher stripped the S_IFIFO type bits (mode & 0o7777) and made a plain file. A concurrent reader of that "FIFO" saw an empty regular file and got a premature EOF.

This breaks bash process substitution read < <(cmd): bash mknod()s a FIFO in $TMPDIR (/tmp/sh-np.XXXXXX) and reads it while the writer subshell is still producing output. Traced end-to-end in a Kandelo guest: the reader's read returned 0 because the temp file was an empty S_IFREG (fstat fifo=false file=true size=0).

It blocked Homebrew in Kandelo: brew config/doctor parse git --version via read < <(printf …) and discover ruby via while read < <(which -a ruby). Both got empty output → "Please update your system Git" and "No Homebrew ruby available for wasm32". Both collapse to this one bug.

What changed

• crate::fifo: registry mapping a FIFO's canonical path → backing pipe index.
• mknod/mknodat(S_IFIFO) register a kernel pipe (PipeBuffer::new_fifo, no endpoints open) instead of creating a regular file.
• open() connects a read or write end to that shared pipe (add_reader/add_writer) → real block/EAGAIN/EOF semantics. Fork inheritance works through the existing negative-host_handle FileType::Pipe OFD path (no new fork logic).
• A FIFO reader with no writer yet blocks (EAGAIN) rather than reporting EOF, until the first writer connects — fixes the reader-opens-before-writer race.
• stat/lstat/newfstatat report S_IFIFO for registered FIFOs (a FIFO has no VFS inode).
• unlink/unlinkat remove the name and free the pipe once idle; FIFO pipes are exempt from is_fully_closed so they persist until unlinked.

Validation

• Process-substitution repro: read a < <(sleep 0.5; printf X); echo [$a] → [X] (was []).
• brew config --no-install-from-api now exits 0 inside Kandelo (previously failed on git-version + portable-ruby). API-mode config now proceeds past both blockers to a separate downstream curl issue.
• 964 kernel unit tests pass, incl. new test_fifo_named_pipe_semantics (EAGAIN-before-writer, data after write, EAGAIN-when-empty-open, EOF after writer close, stat S_IFIFO, unlink).
• No ABI change — reuses FileType::Pipe; no new syscalls or repr(C) structs; abi/snapshot.json unchanged.

🤖 Generated with Claude Code

[brandonpayton]

Absorbed by rebase-merged #967. Main is now 6efb411 with the sealed tree b133bfa018f7e2bc84c4845edf2f945b2f844954, and ABI 41 activation completed successfully in run 29550069085. The absorption audit confirmed this PR\x27s behavior and tests are present in that merged tree, so this source PR is closed and its branch is deleted.