fix(embed): pin stdin to DEVNULL when spawning the daemon

[E-R-Butch]

_detach_popen_kwargs pins stdin on Windows but not on POSIX, so the daemon child inherits the caller's fd 0. When that fd 0 is a socket opened with FD_CLOEXEC (e.g. a TUI/gateway parent that wires its IPC channel onto fds 0-2), the kernel closes it at exec: the child starts with sys.stdin = None and dies in _redirect_stdio_to_log() with

AttributeError: 'NoneType' object has no attribute 'fileno'

The manager then waits out its startup timeout and reports "Daemon failed to start". The sibling helper in hindsight-api-slim/hindsight_api/daemon.py already pins stdin; only this copy had drifted.

Repro (no daemon involved):

import fcntl, os, socket, subprocess, sys
s1, s2 = socket.socketpair()
pid = os.fork()
if pid == 0:
    os.dup2(s1.fileno(), 0)
    fcntl.fcntl(0, fcntl.F_SETFD, fcntl.FD_CLOEXEC)
    subprocess.Popen([sys.executable, "-c", "import sys; print(sys.stdin)"], start_new_session=True)
    os._exit(0)
os.waitpid(pid, 0)  # prints None for the child on current main

Pins stdin on both platforms and adds a regression test (test_detach_popen_kwargs_pins_stdin, fails on main, passes with the fix). Verified on macOS / Python 3.11: a daemon start from a CLOEXEC-fd0 parent comes up healthy.

[nicoloboschi]

Thanks! Rebased onto main and pushed one small review fix: the regression test now indexes kwargs["stdin"] directly so a missing key fails loudly. Full CI (including embed on Linux/Windows) is green apart from the unrelated hermes-compat job (upstream hermes-agent now needs ruamel).