158acfcc

test: the OS oracle answers on Darwin, and the harness spells time and paths through it

a73x   2026-09-03 20:59

test: the OS oracle answers on Darwin, and the harness spells time and paths through it

os_oracle.sh becomes one `case "$(uname)"`: the Linux block moves into the
`*)` arm unchanged and a Darwin arm goes above it, where lsof is the fd
oracle and ps answers the process questions because there is no /proc.
Every spelling in that arm was measured on the port box (macOS 26.6, lsof
4.91) rather than guessed:

  - an unlinked mkstemp carrier keeps its old path in lsof, canonicalized
    under /private and with no marker of its own, so the NAME cannot tell a
    carrier from the suite's own captures under the same /tmp/mux- prefix.
    What can is that the path no longer names anything;
  - a pty master is /dev/ptmx, through openpty(3) as well as posix_openpt;
  - a unix socket answers its bound path and no inode, so this arm matches
    the path inside that pid's own fd table where Linux matches the inode;
  - lsof prints a connected UDP socket as LOCAL->REMOTE, so the arrow and
    what follows is cut before the local-address comparison the Linux arm
    already insisted on;
  - a process that exec'd /bin/sh reports /bin/bash, because Apple ships
    /bin/sh as a stub in front of the bash image — true of a copy of
    /bin/sh too, so it is the image and not the name that differs.

Two new helpers, because a group file must never learn which arm it is on:
now_ms prints epoch milliseconds and real_path canonicalizes, where
`date +%s%N` and `readlink -f` are GNU and a group file that spelled either
would be a Linux group file. Every `date +%s%N` bracket loses its /1000000
to match, and every `readlink -f` becomes real_path. udp_local_bound keeps
the /proc/net/udp hex argument on both OSes and the Darwin arm converts it
in shell arithmetic, which forks nothing — it is polled in a loop.

The three legs that time themselves from INSIDE a mux session cannot call a
helper, so os_oracle.sh also hands out the TEXT of a now_ms through
now_ms_snippet, and the harness's own now_ms is eval'd from that same text
so the two cannot drift. It stays a per-OS `date`/`gdate` rather than an
interpreter because the throughput gate's tightest ceiling is 10ms around a
6ms leg, where a python3 start measured 21.6ms would be most of the reading.

oracle_selftest moves to test/oracle_selftest.sh, sourced by e2e_lib.sh and
runnable on its own (`sh test/oracle_selftest.sh`, answering ORACLE_OK) so
that a new OS arm can be graded on a box where nothing else in the harness
works yet. Its pid_exe assertion moves off the subject shell and onto that
shell's `sleep` child, which is an ordinary binary on both OSes — asking it
about a shell would have graded Apple's stub and not the helper.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SakwJEwD9dXBoRP5kWbemW