Closes #3263

Three independent fixes for the custom endpoint experience with Ollama, LM Studio, and other local/custom OpenAI-compatible servers.

Fix 1 — Surface connection errors (silent validation failure)

File: hermes_cli/model_switch.py

The except Exception block in Step 5 was swallowing all validation errors and returning accepted: True, so switching to an unreachable http://localhost:11434/v1 would report "Switched successfully" even when Ollama wasn't running.

Before:

> /model llama3
Switched to llama3 (custom)
[chat proceeds to fail silently]

After:

> /model llama3
Error: Could not reach custom endpoint: Connection refused

The fix detects connection-class errors (refused, timeout, 401, 403, 404) specifically on custom/localhost targets. Cloud providers keep the previous lenient behaviour — a temporary OpenRouter outage should not block switching to a model that's likely fine.

Fix 2 — Pre-fill base_url and model from config.yaml

File: hermes_cli/main.py → _model_flow_custom()

The interactive setup flow read OPENAI_BASE_URL from env but ignored model.base_url and model.default in config.yaml, presenting blank prompts every time even after a successful previous setup.

Before:

Custom OpenAI-compatible endpoint configuration:

API base URL [e.g. https://api.example.com/v1]: _
API key [optional]: _
Model name (e.g. gpt-4, llama-3-70b): _

After:

Custom OpenAI-compatible endpoint configuration:
  Current URL: http://localhost:11434/v1
  Current model: llama3

  Press Enter to keep the current value shown in [brackets].

API base URL [http://localhost:11434/v1]: _
API key [optional, press Enter to skip]: _
Model name [llama3]: _

Empty input → keeps the current value. The config is read from config.yaml first, with env vars as fallback.

Fix 3 — config.yaml custom endpoint survives OPENROUTER_API_KEY in .env

File: hermes_cli/runtime_provider.py

When requested_norm was "auto" (normal CLI startup) and cfg_provider was "custom", use_config_base_url was not set. The base_url resolution chain then fell through to OPENROUTER_BASE_URL, flipping the active provider back to OpenRouter on every restart.

The existing #1760 fix only handled requested_norm == "custom" explicitly, missing the "auto" path used at startup.

Fix: added elif cfg_provider == "custom" inside the "auto" block so that an explicitly saved custom endpoint always wins, regardless of what env vars are present.

Tests

8 new tests in tests/test_model_switcher_custom_3263.py:

44 total pass (36 existing + 8 new).