The registry's Exec returned an error wrapping ExitStatus(127) when a
command wasn't registered, but never wrote the diagnostic anywhere.
Callers like time and env that dispatch through the runner saw status
127 with no message, so users got a silent failure (and the matching
tests failed expecting "command not found" on stderr).

Assisted-by: Claude Opus 4.7 via Claude Code
Signed-off-by: Xe Iaso <me@xeiaso.net>

The previous pipe layout shared one stdin pipe between term.Terminal
(prompt) and the foreground command. That conflated two lifetimes:
closing the pipe to deliver EOF to a wasm guest on Ctrl-D would also
take down the prompt, and not closing it left REPLs unable to exit
their inner read loop.

Split it into two pipes. The prompt pipe is long-lived — term.Terminal
reads it for the entire SSH session. The command pipe is rotated each
sh.Run: a fresh os.Pipe before the run, handed to the runner via
interp.StdIO, closed and dropped after.

The SSH-input pump now routes bytes by mode (gated on the existing
commandActive atomic): prompt pipe while idle, command pipe while a
command is foreground. While in command mode, it scans for VEOF
(Ctrl-D, 0x04) — the same byte a kernel PTY recognises — and on hit,
flushes anything before it, then closes the command's stdin so the
guest's next fd_read returns 0. A small mutex guards the rotating
write handle so the pump can race-safely close it from its goroutine
while the main loop swaps in the next pipe.

Assisted-by: Claude Opus 4.7 via Claude Code
Signed-off-by: Xe Iaso <me@xeiaso.net>

Bare \`python3\` under sophia would slurp stdin to EOF before doing
anything — Python's non-TTY default treats stdin as a script. Inject
-i when no args are given so it eval-prints line by line, and -S to
skip site.py: under -i, site appends cwd ('/') to sys.path and
listdir-walks '/' trying to import readline, which fails on the s3fs
mount and takes the REPL down before the first prompt.

Also set PYTHONUNBUFFERED=1 so the interpreter flushes after each
line instead of waiting for its block buffer to fill. The pipe-backed
stdio sophia now provides is line-oriented from the user's side; an
unflushed Python buffer makes the REPL look frozen.

Assisted-by: Claude Opus 4.7 via Claude Code
Signed-off-by: Xe Iaso <me@xeiaso.net>

Previously the SSH session handed the kefka shell a strings.NewReader
for stdin and a *term.Terminal for stdout/stderr. wazero only takes
its fast *os.File path when host stdio is genuinely *os.File; the
term-backed handles fell back to a treat-stdio-as-block-device path
that broke wasi-libc isatty detection inside python.wasm and qjs.wasm,
so REPLs would not start a real interactive loop.

Replace that with three os.Pipe() pairs threaded through interp.StdIO,
plus a goroutine that pumps SSH bytes into the shared stdin pipe with
software line discipline that a kernel PTY would normally do:

- ICRNL: translate Enter (\r on a raw SSH channel) to \n, so
line-buffered WASI readers like Python's fgets recognize line
boundaries. term.Terminal accepts either, so the prompt is
unaffected.
- ECHO: while a foreground command is active (gated by an
atomic.Bool flipped around sh.Run), echo typed bytes back to the
client so REPLs aren't typing blind. term.Terminal already
handles its own echo during the prompt.

stdout/stderr drain into the terminal via io.Copy goroutines so its
writeWithCRLF still does \n -> \r\n on the way back out.

Assisted-by: Claude Opus 4.7 via Claude Code
Signed-off-by: Xe Iaso <me@xeiaso.net>

Signed-off-by: Xe Iaso <me@xeiaso.net>