fix: provider-aware model routing for step-memory lenses

[kypeli]

The Flash step summarizer (and, latently, the chunk-capsule lens) forced every explicitly configured model through get_google_llm(), which hardcodes provider=GOOGLE. A non-Gemini model ID such as "tensorx/deepseek/deepseek-v4-flash" was therefore sent verbatim to Google's Generative Language API, producing a terminal, non-retryable 404 per step:

VisualStepSummarizer: Error generating summary for step 1:
Error calling model 'tensorx/deepseek/deepseek-v4-flash'
(Not Found): 404 Not Found. {'message': '', 'status': 'Not Found'}

The error surfaced only at call time because construction errors were silently swallowed (the old try/except re-fell back to get_google_llm()), and 404 classifies as BAD_REQUEST (non-retryable, no fallback) in artemis.llm.reliability. Other nodes resolved the same model ID correctly: they route through _resolve_endpoint(), which maps the configured provider — "custom" nodes ride the OpenAI-compatible gateway (OPENAI_BASE_URL) where namespaced IDs are valid.

Add get_lens_llm(ctx, model_name, provider) to artemis.services.llm: a provider-aware, raw-model factory for step-memory lens calls. It returns the raw chat model (deliberately NOT wrapped in RobustChatModelWrapper) so the lenses keep their own bounded retry loops, 25s asyncio timeouts, and explicit "lens:*" usage metering. Provider resolution order:

1. Explicit "provider" knob — validated via ModelProvider.from_string, so typos fail fast with the standard "Unknown LLM provider" error.
2. Gemini model names (bare, or google/gemini-prefixed) route to GOOGLE, keeping every pre-existing config byte-for-byte compatible.
3. Anything else inherits the LLM config's summarizer node provider (which itself inherits the global "default" block). Namespaced model IDs pass through verbatim — never parsed as provider prefixes.
4. Resolution failures default to CUSTOM, matching how utils nodes behave when no config is live.

Also emit a loud warning when a non-Gemini model resolves to the Google Gemini API — the historical misroute produced no signal anywhere in the stack until the first 404.

Changes:

• artemis/services/llm.py: add get_lens_llm, _inherit_lens_provider, and _warn_if_non_google_routes_to_google.
• artemis/config/agent.py: add optional "provider" field (validated at load) to StepSummarizerConfig and MemoryChunkingConfig; expand model field descriptions with the routing rules. Old configs parse unchanged (fields default to None = auto-resolve).
• artemis/agents/flash/summarizer.py: replace the get_google_llm / dead get_llm(is_utils=True) branches with a single get_lens_llm call; surface construction errors instead of silently swapping providers.
• artemis/agents/flash/runner.py and artemis/memory/init.py: forward step_summarizer_cfg.provider (covers the Pro profile's shared step-memory service too).
• artemis/memory/chunking.py: same fix for StepCapsuleLens — accept model_provider/fallback_model_provider and route through get_lens_llm. The capsule fallback now resolves the summarizer fallback block's OWN provider (like every other node's fallback via _resolve_endpoint(use_fallback=True)) instead of being restricted to google-only or inheriting the primary's provider; a Gemini model under a non-Google knob still auto-routes to GOOGLE.

Tests:

• New tests/unit/services/test_llm_lens_routing.py: the full routing matrix (Gemini shortcut, google/gemini prefixes, explicit knob precedence, unknown-provider failure, summarizer-node inheritance, default-config fallback, CUSTOM-on-failure, non-Gemini→GOOGLE warning) plus an end-to-end assertion that the previously-404ing TensorX model builds a ChatOpenAI bound to the configured gateway base URL.
• test_history_chunking.py: fallback resolution now returns (model, provider) and is provider-aware; new precedence tests for the fallback block's own provider vs. the chunking "provider" knob.
• test_memory_config.py: the new provider knobs parse from config and reject unknown providers.
• test_flash_runner.py: FlashRunner forwards step_summarizer provider.

Closes #167

[kypeli]

Disclaimer: LLMs were used to create this merge request, but I did review it manually, tested and verified the changes locally.