AI Assistant Updates: Model, UI, and Fixes

Regenerated search tool provider embeddings, restoring 1142 provider examples after a chunk-size regression that reduced coverage using the current EmbeddingService.

Fixed two test failures: one involved updating hard-coded DDL in a test to include the missing tool_call_id column, and the other updated an assertion related to CPU memory arena settings to match intentional base model behavior.

Implemented session heartbeat and auto-commit for the brain UI. The session heartbeat polls for status and redirects to login upon session expiry or vault lock.

When saving a provider, any text remaining in the model input is now automatically promoted to the models list, preventing data loss.

Added logic to redirect the client to the login screen when the server session becomes invalid, tied to the new client heartbeat polling /auth/status.

Updated default model references across documentation, UI placeholders, and test fixtures from qwen3 to gemma4:31b without changing runtime behavior.

• Regenerated 1142 search tool provider embeddings to correct chunk-size regression.

• Fixed test failure in test_tool_call_service by adding tool_call_id to test DDL.

• Brain UI now features session heartbeat to detect expired or locked sessions.

• Provider form auto-commit promotes remaining model input when saving a provider.

• Client is redirected to login when server session validation fails.

• Default model references were updated to gemma4:31b across the codebase.