Registry disruption and model poisoning via encoded names

Model-name validation failure in MLflow before 2.11.3 allows remote authenticated users to poison registry data via URL encoding. The Model Registry creation path does not canonicalize encoded names before enforcing uniqueness, so encoded and decoded forms create distinct records that resolve ambiguously. Model-creation access is required; resulting collisions can redirect later model selection to attacker-controlled data or interrupt access to the intended model.