Limit local catalog to current OpenZero models

[ResearchForumOnline]

What changed

• Limit OpenZero model APIs and the Super Panel local selector to the three current personal releases: openzerogemma:latest, zero-qwen3-q5:latest, and zero-qwen3-f16:latest.
• Filter the custom GGUF inventory to the matching three release files.
• Stop the UI from re-inserting a deleted saved local alias as Resolved Local.
• Align the public model documentation and Tab Pilot page with the working Qwen aliases.
• Add a regression contract covering the three-model catalog and stale-alias behavior.

Why

The panel was populated from every Ollama alias installed on the node, and ZeroThink/OpenZero could continue surfacing aliases whose source GGUF files or catalog records had already been removed. The saved-model fallback could then add one of those stale aliases back into the selector.

Impact

Operators see only the three current OpenZero releases in the local model lane. Cloud providers remain available, older documented Qwen spellings continue to normalize to the working aliases, and unrelated Ollama models are not destructively removed by an update.

Checks

• 21 OpenZero contract tests passed.
• brain/app.py compiled successfully.
• git diff --check passed.
• The same runtime changes are already live and verified through /api/models.
• The published release ZIP passed archive, SHA-256, internal manifest, and independent HTTPS download verification.