Hook into CursorManager::timeoutCursors to check for memory load shedding at the same 4s frequency that we check for timeouts. The challenges here revolve around the lack of opCtx:

• No opCtx in the sweep → kill the cursor directly (dispose / executor->markAsKilled(QueryMemoryLimitExceeded)), like timeout does, rather than opCtx->markKilled.
• dt = sweep cadence (clientCursorMonitorFrequencySecs, ~4 s) and shedProbability caps dt at 1 s, so idle cursors would under-shed unless we give this path its own interval/cap; seed lastShedEvalTime when the cursor goes idle.
• The roll runs under the partition lock — keep the disabled/below-low-mark short-circuit so it's ~free when healthy; dispose already defers outside the lock.
• mongos idle cursors live in ClusterCursorManager (the SERVER-131578 router gap), separate from this.