fix(models): preserve Ollama model:tag colons in parse_model_input

[hashbender]

Problem

When ACP clients (like AionUI) send custom:glm-5.2:cloud to set_session_model, parse_model_input splits at the second colon, treating the middle part (glm-5.2) as a named custom provider name. For Ollama models that use colons in their tags (e.g. glm-5.2:cloud, nomic-embed-text:latest, kimi-k2:1t-cloud), this truncates the model id to just the tag suffix (cloud instead of glm-5.2:cloud), causing HTTP 404 errors from the inference endpoint.

The existing triple-syntax guard (custom:name:model → ("custom:name", "model")) had no way to distinguish between:

• custom:ollama:qwen3 — provider custom:ollama, model qwen3 (intended triple syntax)
• custom:glm-5.2:cloud — provider custom, model glm-5.2:cloud (Ollama tag, NOT triple syntax)

Solution

Add _is_registered_custom_provider() that checks whether the middle part matches a configured custom provider in config.yaml (via get_compatible_custom_providers()). The triple split only happens when the middle part is a registered custom provider name. Otherwise, the full string after custom: is treated as the model id.

# Before (truncates Ollama tags):
if custom_name and actual_model:
    return (f"custom:{custom_name}", actual_model)

# After (only splits for registered custom providers):
if custom_name and actual_model and _is_registered_custom_provider(custom_name):
    return (f"custom:{custom_name}", actual_model)

Testing

• Updated test_custom_triple_syntax and test_custom_triple_spaces to mock _is_registered_custom_provider (since local-server/my-server are not registered providers in the test env)
• Added test_custom_triple_unregistered_falls_back — verifies custom:glm-5.2:cloud with a mocked False returns the full model id
• Added test_custom_ollama_tag_not_truncated — regression test without mocks, verifies real-world Ollama model names pass through intact

Verified manually:

parse_model_input("custom:glm-5.2:cloud", "custom")    → ("custom", "glm-5.2:cloud") ✅
parse_model_input("custom:ollama:qwen3", "custom")      → ("custom:ollama", "qwen3")  ✅
parse_model_input("custom:nomic-embed-text:latest", ..) → ("custom", "nomic-embed-text:latest") ✅
parse_model_input("custom:gpt-4o", "custom")            → ("custom", "gpt-4o")         ✅

Scope

• Only affects the triple-syntax path (custom:name:model). Single-colon path (custom:model) is unchanged.
• No new dependencies. _is_registered_custom_provider uses get_compatible_custom_providers which already exists in hermes_cli/config.py.
• Related: fix(acp): split custom:name/model and provider/model model names (#53545) NousResearch/hermes-agent#53558 addresses a similar parsing issue for / (slash) syntax in ACP model names. This PR is complementary (different delimiter, same root concern).

---

Mirror-of: NousResearch#56089
NousResearch#56089