feat(llm): per-node local/edge VLM endpoints + JSONC URL fix + provider-surface tests

[RanjithRagavan]

Summary

Unblocks the On-Device Lightweight VLMs roadmap item by making local/edge OpenAI-compatible endpoints (Ollama, llama.cpp-server, vLLM, LM Studio) configurable per agent node — today only the global OPENAI_BASE_URL env var works, forcing every node onto one endpoint.

Closes #131

Relation to #128: complementary, not overlapping. #128 adds admin-UI/env credential plumbing (CUSTOM_LLM_API_KEY, Settings.get_api_key) and validate_provider branches. This PR adds the per-node config→endpoint passthrough in _resolve_endpoint(), repairs the local-ollama preset, fixes a latent JSONC parser bug, and adds provider-surface tests. Happy to rebase whichever lands second.

Changes

Endpoint plumbing

• artemis/config/llm.py: LLM gains optional api_base, api_key, max_tokens, timeout_seconds, is_multimodal (types match ModelEndpoint); validate_provider() no longer requires cloud keys for ollama / vllm / custom
• artemis/services/llm.py: _resolve_endpoint() forwards all five fields into ModelEndpoint, for primary and fallback nodes

Preset + parser fixes

• local-ollama preset (both config copies): provider: "ollama", model: "qwen2.5vl:7b", api_base: "http://localhost:11434/v1" — boots with zero cloud keys
• artemis/utils/file.py: strip_json_comments() rewritten as a string-aware scanner — the old regex stripped // inside quoted strings, corrupting any config containing a URL value (blocking bug for any config-level endpoint fix)
• .env.example documents local-endpoint variables

Tests

• New tests/unit/test_llm_router.py: 22 tests — provider alias parsing, mocked ChatOpenAI construction for OLLAMA/VLLM/CUSTOM, _resolve_endpoint passthrough/defaults/fallback, local-provider validation, JSONC string safety, preset regression

Docs

• docs/design/on-device-vlm.md: design doc for the on-device VLM roadmap item (deployment shapes, endpoint schema with hybrid local-flash + cloud-pro example, perception-pipeline next steps, privacy model, latency measurement plan); roadmap bullets in both READMEs link to it

Drive-by CI fix

• playground/backend_manager: auto-fixed 11 pre-existing ruff 0.16.3 lint errors (datetime.UTC alias style) that fail the Lint Python step under the ruff version CI resolves today — verified failing on a clean checkout of main

Test plan

• pytest tests/unit/test_llm_router.py tests/unit/test_llm_reliability.py tests/unit/test_llm_structured.py tests/unit/test_llm_gateway.py — 44 passed
• Full pytest tests/unit — failure set identical before/after this branch (pre-existing, environment-related e.g. missing GOOGLE_API_KEY)
• ruff check + ruff format --check clean on the whole tree
• LLM(provider="ollama", model="qwen2.5vl:7b", api_base="http://localhost:11434/v1") validates with no cloud keys set

Example: hybrid on-device setup now possible

"default": { "provider": "google", "model": "gemini-2.5-pro" },
"nodes": {
  "explorer": {
    "provider": "ollama",
    "model": "qwen2.5vl:7b",
    "api_base": "http://localhost:11434/v1",
    "is_multimodal": true
  }
}

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[RanjithRagavan]

@google-cla I signed it!

[RanjithRagavan]

@google-cla I signed it!