Summary

Fixes #4172. Supersedes #4174 (different root cause than described there).

The actual bug: _model_flow_custom() saved model.provider: "custom" and model.base_url to disk via its own cfg = load_config() / save_config(cfg) cycle — but never updated the setup wizard's in-memory config dict. The wizard's final save_config(config) at setup.py line 1752 then overwrote the custom settings with the stale default (model: "anthropic/claude-opus-4.6" string).

After this, resolve_requested_provider() saw no model.provider in config, returned "auto", and the auto-detection chain picked up stale OPENROUTER_API_KEY from .env → traffic routed to OpenRouter.

Every other provider in the setup wizard uses _set_model_provider(config, ...) which correctly updates the wizard's dict. The custom flow was the only one using a separate load_config().

Fix: After saving to disk, also mutate the caller's config dict so the wizard's final save preserves model.provider="custom" and the base_url. Both the model-name and no-model-name branches are covered.

What changed

• hermes_cli/main.py — _model_flow_custom() now syncs the passed config dict after its own save (+15 lines)
• tests/hermes_cli/test_setup.py — Updated existing test to include the wizard's final save_config(config) call (the step that was previously untested), plus two new regression tests covering the Ollama and no-model-name scenarios

Test plan

• 5002 tests pass (2 pre-existing failures unrelated to this change — broken by refactor: make config.yaml the single source of truth for endpoint URLs #4165's OPENAI_BASE_URL removal)
• E2E: fresh install → custom endpoint → wizard's save_config → reload → resolve_requested_provider() returns "custom"
• E2E: stale OPENROUTER_API_KEY + custom endpoint → custom endpoint wins, base_url points to localhost
• E2E: no model name entered → model.provider and model.base_url still survive