Preserve context_length when switching desktop models

[LeonSGP43]

Summary

• preserve an explicit override when the desktop model picker switches the main model
• keep the existing provider/base_url reconciliation logic unchanged
• add a regression test covering a fresh Ollama session with a custom context length

Testing

• uv run --frozen --extra dev pytest tests/hermes_cli/test_web_server.py -k set_model_assignment_preserves_context_length_override
• uv run --frozen ruff check hermes_cli/web_server.py tests/hermes_cli/test_web_server.py
• git diff --check

Closes #41944

[teknium1]

Thanks for identifying the desktop model-assignment path. The premise is still valid on current main: hermes_cli/web_server.py:1141 unconditionally removes model.context_length, despite the documented highest-priority config override in website/docs/integrations/providers.md:1141 and resolver behavior in agent/model_metadata.py:2051-2053.

Problems

• The current direct helper contract at tests/hermes_cli/test_web_server.py:3162-3168 still asserts that context_length is removed. The added endpoint regression does not replace that assertion, so the updated production behavior would leave contradictory test expectations.
• GitHub currently reports the branch as conflicting. The helper has since grown endpoint-key handling at hermes_cli/web_server.py:1130-1140; preserve that logic while applying the context override change.

Suggested changes

• Update the direct helper test to assert retention of an explicit context_length, alongside the new endpoint-level regression.

Automated hermes-sweeper review.