A test-time continual model merging framework that leverages low-rank mixture-of-experts and null-space constrained gating to adapt dynamically to new tasks while preserving prior knowledge, all without access to training data.