verify-backends ignores run_acq's exit status when classifying USAi reachability, so a failed exec is recorded as pass #547

verify_backend runs the USAi reachability probe through run_acq, then classifies the
result purely by string-matching the captured output, and never looks at the command's exit
status:

run_acq "exec usai-reach" --backend "$backend" exec "$name" -- sh -c \
  "curl -sS -o /dev/null -w '%{http_code}' --max-time 15 $USAI_MODELS_URL"
local ustatus="$LAST_OUT"
case "$ustatus" in
  000|*"resolve"*|*"curl:"*|"")
    fail "${backend}: USAi host not reachable from guest (egress/net-rule?)"
    ...
  *)
    pass "${backend}: USAi host reachable from guest (HTTP $ustatus; auth rewritten on wire)"
    ;;
esac

scripts/verify-backends:612-623 on main (e0b1143b) — 610-623 including the run_acq
invocation above.

run_acq does return the command's status (return "$_rc"), so the information is
available; the caller discards it. Any output that is non-empty and happens not to contain
the substrings resolve or curl: lands in the *) branch and is recorded as pass —
including a backend error, a shell error, or a curl diagnostic whose wording changes
upstream. The pass message then interpolates that text as if it were an HTTP status.

Reproduced in isolation: a stand-in command printing noisy-error and exiting 7 was
classified pass.

This is the exact false-green class the live harness exists to prevent, in the one check
that proves egress actually works.

Suggested fix: check run_acq's return value first, and require ustatus to match a
three-digit HTTP status ([0-9][0-9][0-9]) before treating it as reachable — anything else
is a failure, not a pass.

One caveat on the status source, if this is fixed on top of
GSA-TTS/agentic-coding-quickstart#474 rather than on main: on main, run_acq's
verbose branch reads PIPESTATUS[0] after the enclosing LAST_OUT=$(... | tee /dev/stderr)
assignment, which reports the assignment's status (0) rather than the command's. So under
-v on main, the return value is not currently trustworthy either, and a fix that checks
it needs that corrected too. PR #474 already rewrites this path (_run_captured runs the
command directly and preserves the real status), so on that branch the rc is sound and only
the caller needs changing.

Provenance: surfaced during review of GSA-TTS/agentic-coding-quickstart#474, where it
was initially filed as a blocking finding against that PR. It is not that PR's defect — the
block is byte-identical to main and does not appear in its diff. Filing it here so it is
tracked against the right target.