The keep-going Stop hook only existed for an active plan. Now the user can say
"keep going until X" at any moment: `commandments until "<condition>"` records it
in the session, and while any condition stands the UntilReminder Stop hook holds
every stop and sends the agent back to VERIFY it rather than assume it.

Conditions stack, each with its own number. `until met <n>` strikes a verified one
off (the gate lifts when none remain), `until stuck` releases ONE stop for a
genuinely blocked agent while keeping the condition in force, `until clear` drops
the gate — the user's call, never an escape hatch. Loop-safe: ten held stops with
no progress and the gate releases itself and tells the agent to report back;
meeting a condition resets that count.

The human gets a handle too: a `/until <condition>` slash command published into
`.claude/commands/` on sync, alongside the `commandments-until` skill that teaches
the discipline.

PlanMarker and the new gate wanted the same "value lines + separator + explanation"
marker format, so that is now one MarkerFile both compose.