d6888aa5

test: e2e — quality batch: full SOCK17 teardown, trap backstops, fixture hardening, honest comments

a73x   2026-08-12 04:25

test: e2e — quality batch: full SOCK17 teardown, trap backstops, fixture hardening, honest comments

From the Task 7 quality review. No scenario changes its coverage; every
item either closes a gap in a teardown, hardens the fixture, or makes a
comment say what is actually true.

SOCK17's teardown asserted only the stop's exit code. It now makes the
same three assertions the SOCK16 leg does — exit 0, the daemon's own
`muxd: stopped` line, and the socket gone — because an exit code says the
command returned, not that it did the job, and a stop that exited 0 while
leaving a socket behind was invisible here.

$MUX_KEY_FILE joins $SHELL and the XDG homes at the top rather than being
cleared inside the M14 block: `muxd endpoint` reads it BEFORE the default
key path, so an operator with one exported would have changed which key
the handoff announced, and the scenarios would have gone on passing
against the wrong key until one of them did not. Same argument 04de1f6
made for $SHELL, same place, so a later scenario cannot forget it.

The three backgrounded clients are now tracked pids the trap can kill.
Their `timeout` remains the primary bound; this only shortens the wait on
a run that failed between a launch and its wait. (a)'s launch also takes
the set +e shape (d) and (e) already used, so all three read alike.

The ssh shim refuses fewer than two arguments with 97. A fixture whose
silence cannot be told from success is worth two shell words to rule out,
and the guard sits BEFORE the pidlog append, so a malformed call cannot
inflate the ssh counts the scenarios assert on. Checked both ways: 0 and
1 arguments exit 97 and log nothing, 3 arguments run and log one line. A
comment now also states the contract this depends on — the client spawns
`ssh HOST CMD...` as separate words — and what a future client that
quoted the remote command would do here (fail at 127, loudly).

Two comments were wrong or incomplete rather than merely thin:

(d)'s ceiling now says what it can SEE. Past about six seconds the
script's detach lands inside the dial, waitReady answers it as an abort,
no fallback line is printed at all, and the overrun is reported by the
wait_for with its own message. This bound owns the band between the
budget and that; the wait_for is the backstop beyond it. 10000 stands.

(a)'s baseline now names the real guarantee — openHandoff kills ssh the
moment QUIC is ready, before Transport.open returns and therefore before
the session can emit anything — so a reader who moves that kill knows
they are taking the scenario's premise with them. Its failure message was
also inaccurate: it accused the product of an ordering the product does
not get wrong. When an attach outruns the 6s script both markers are
buffered and delivered together, the kill still precedes both, and what
is lost is the scenario's ability to WITNESS the ordering. It now says
that instead.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>