Say how to finish a failed teardown, and record the probe's real history

A failed --remove told the user to run it again, but that advice could
never work: rm -f has usually taken the marked stub by the time the
failure is judged, and a rerun that finds nothing it owns succeeds
without touching the mise environment it was asked to finish removing.
Spell out the three commands that complete the job by hand instead.

Also correct the story the probe comment told: chat never lost
--oneshot in v0.20 -- no released Hermes defined it there. It lived at
the top level until v0.21 added chat's own, so the old probe was keyed
to a flag no release ever carried under chat, and every install read
as not ready. Recorded straight so a future hermes-desktop bump to
v0.21+, which would make the old probe pass on the desktop path alone,
cannot read as the fix.

Findings from omarchybot's review (Opus 5, with Codex at xhigh).

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

test/shell.d/hermes-cli-test.sh

@@ -499,8 +499,8 @@ chmod +x "$ready_home/.hermes/hermes-agent/venv/bin/hermes"
 run_ready_check || fail "--check accepts the app's wrapper once it runs"
 pass "readiness runs the app's command rather than trusting its marker"
 
-# A release whose help lists only the retired --oneshot marker cannot run the
-# seeded --tui --query session omarchy-agent starts, so it is not ready.
+# A release whose help lists only the old probe's --oneshot marker cannot run
+# the seeded --tui --query session omarchy-agent starts, so it is not ready.
 cat >"$ready_home/.hermes/hermes-agent/venv/bin/hermes" <<'SH'
 #!/bin/bash
 if [[ ${1:-} == "chat" && ${2:-} == "--help" ]]; then