Follow-up to #14428.

Problem. escaping_weight_references reads a Sentence Transformers folder's modules.json and flags module folders outside the model. Before the escape check it skipped paths that strip("./\\") empties, meant for the model root ("", ., ./). But .., ../.., / and \ are also only dots and slashes, so they were skipped too, and a managed account's model folder could name a module folder above it (its account root, or the Studio root) without the load, export or training gate noticing.

Change. The escape check now runs first; the self-folder skip comes after it. A module at the model root is still accepted, so ordinary Sentence Transformers folders load as before. Owner accounts are unaffected (the gate only runs for managed accounts), and notebook loading in unsloth is unchanged.

Testing.

• test_account_path_confinement.py: 38 passed. The 6 new escape cases (.., ../.., /, \, ./.., ..\..) fail on main; the root cases ("", ., ./) pass on both.
• Neighbouring suites (test_embedding_model_security_gate.py, test_export_decision_backend.py, test_decision_training_route.py, test_export_absolute_paths.py): 134 passed. One test_export_absolute_paths.py case fails the same way on main in this environment (core.inference.diffusion_lora not importable).