fix: address PR review feedback on LLM channel config

[Copilot]

Four bugs identified in review of the multi-channel LLM config PR, covering false validation errors, incorrect model prefix canonicalization, swallowed error semantics, and a keyboard UX regression.

src/config.py — false positives in validate_structured()

Direct-env LiteLLM providers (groq/*, bedrock/*, etc.) were being flagged as "not in channel model list" for primary, fallback, and vision models. Added _uses_direct_env_provider() guard to all three checks.

api/v1/endpoints/system_config.py — error status on test_llm_channel

All exceptions were returning 500 internal_error, making it impossible for the client to distinguish bad input from a backend fault. ValueError/TypeError now return 422 validation_error; unexpected errors remain 500.

LLMChannelEditor.tsx — double-prefix on alias provider names

normalizeModelForRuntime() would produce openai/vertexai/gemini-pro instead of vertex_ai/gemini-pro for models with alias prefixes not in KNOWN_MODEL_PREFIXES. Added PROTOCOL_ALIASES canonicalization (vertexai→vertex_ai, claude→anthropic, google→gemini, etc.) applied before the prefix lookup:

const PROTOCOL_ALIASES: Record<string, string> = {
  vertexai: 'vertex_ai',
  vertex: 'vertex_ai',
  claude: 'anthropic',
  google: 'gemini',
  openai_compatible: 'openai',
  openai_compat: 'openai',
};
// alias prefix is canonicalized before KNOWN_MODEL_PREFIXES check

LLMChannelEditor.tsx — Space key scrolls page on channel header

onKeyDown was missing e.preventDefault() for Space, causing the page to scroll when keyboard-navigating channel rows. Fixed.

---

📍 Connect Copilot coding agent with Jira, Azure Boards or Linear to delegate work to Copilot in one click without leaving your project management tool.