llm_factory: cache parsed llm_config per profile

[neoneye]

Summary

Production logs were full of repeated INFO/WARNING noise at every Luigi task start:

INFO  planexe_config        - Optional file '.env' not found …
WARN  planexe_llmconfig     - Environment variable 'OPENAI_API_KEY' not found.
WARN  planexe_llmconfig     - Environment variable 'DASHSCOPE_API_KEY' not found.
WARN  planexe_llmconfig     - Environment variable 'DEEPSEEK_API_KEY' not found.
INFO  planexe_llmconfig     - Applied PLANEXE_LLM_CONFIG_WHITELISTED_CLASSES=…

…because every LLM lookup was re-reading the JSON, re-substituting env vars, and re-applying the class whitelist.

Now _load_llm_config caches the parsed PlanExeLLMConfig in a module-level dict keyed by resolved ModelProfileEnum. Subsequent calls hit the cache. Pricing and Anthropic usage-hook side effects move into the cache-miss branch — same end state, just no longer redone per call.

Latent mutation bug fixed in passing

get_llm was mutating config["arguments"] in place when merging caller kwargs. Pre-cache that was harmless (the dict was rebuilt every call). With caching it would leak state across calls. Shallow-copy the dict before update().

Not in this PR

The OPENAI_API_KEY / DASHSCOPE_API_KEY / DEEPSEEK_API_KEY warnings still fire on the first load even though those entries are dropped by the OpenRouter class whitelist. Fixing that needs filter_by_whitelisted_classes to run before substitute_env_vars in PlanExeLLMConfig.load. Saved for a separate PR.

Test plan

• Boot worker_plan, run a plan, confirm the Optional file '.env' not found / Environment variable …_API_KEY not found lines now appear at most once per profile per worker (instead of per task).
• get_llm("foo", max_tokens=8192) followed by get_llm("foo") — the second call must NOT see max_tokens=8192 leaked in.

🤖 Generated with Claude Code