fix(vertex): prefix google publisher for Gemini models

[kyssta-exe]

Fixes #56778.\n\n## Summary\n- normalize bare Gemini/Gemma model IDs for the Vertex AI provider to the required google/ publisher-qualified form\n- preserve already-qualified and custom publisher model IDs\n\n## Tests\n- scripts/run_tests.sh tests/hermes_cli/test_vertex_provider.py tests/hermes_cli/test_model_normalize.py

[harjothkhara]

Recommendation: treat #56806 as the stronger lane for #56778, and let #56805 either adopt this normalization shape or close as a duplicate.

Cluster disposition:

• #56806: strongest / merge-candidate lane. It fixes bare Gemini/Gemma Vertex IDs while preserving already publisher-qualified IDs (custom-publisher/model-id) and unknown private model IDs, and it covers the google-vertex alias.
• #56805: duplicate / needs-improvement. Its current Vertex block unconditionally prefixes every non-google/ value after stripping vertex/, so custom-publisher/model-id becomes google/custom-publisher/model-id.

That distinction matters because the Vertex error in #56778 says the endpoint expects <publisher>/<model>, not that every publisher-qualified model should be rewritten to Google's publisher namespace. Preserving slash-qualified IDs keeps the fix bounded to the reported bare Google-model case without breaking other Vertex publisher namespaces.

Proof I checked:

• exact duplicate search for 56778 found #56805 and #56806; broader Vertex AI google publisher prefix Gemini search found historical #16010/#21480 only.
• On #56805: normalize_model_for_provider("custom-publisher/model-id", "vertex") returns google/custom-publisher/model-id.
• On this branch: the same call returns custom-publisher/model-id; bare gemini-3.1-flash-lite and alias google-vertex both normalize to google/gemini-3.1-flash-lite.
• scripts/run_tests.sh tests/hermes_cli/test_vertex_provider.py tests/hermes_cli/test_model_normalize.py -> 100 passed.

[kyssta-exe]

Stale — 12-13 days without merge activity.