Drop stale llm_knowledge_base docs + dead OpenAI consumer

[lacraig2]

What

Removes docs/llm_knowledge_base/{playbook,plugins}.md and the only code that consumed them (src/penguin/llm.py + the dead GraphSearch.select_best_config_llm path).

Why

The llm_knowledge_base docs have drifted out of sync with the tool and are actively misleading to an LLM agent that reads them:

• Wrong output filenames — they say console.txt, env_missing.txt, pseudofiles_failures.txt, pseudofiles_modeled.txt, base/fs.tar. The real artifacts are console.log, env_missing.yaml, pseudofiles_failures.yaml, pseudofiles_modeled.yaml, base/fs.tar.gz (verified in pyplugins/analysis/env.py, pyplugins/hyperfile/pseudofile_tracker.py).
• Nonexistent script — they tell the reader to "customize makeuboot.py"; no such file exists in the tree.
• Imaginary tool interface — "invoke your add_pseudofile tool", "your environment_variable tool". Penguin has no such tool API; it is driven by editing the YAML config and re-running penguin run.

The docs' only consumer is src/penguin/llm.py — an OpenAI gpt-4o Assistants/vector-store integration reached via GraphSearch.select_best_config_llm(). That path is dead and broken:

• its sole caller is commented out (graphs.py);
• upload_knowledge_files() references an undefined self.KNOWLEDGE_DIR;
• it uses deprecated openai.beta Assistants/vector-store APIs;
• openai is not a declared dependency.

Changes

• rm docs/llm_knowledge_base/
• rm src/penguin/llm.py
• graphs.py: drop the from . import llm import, the dead select_best_config_llm method, and its commented-out caller. (UUID import retained — still used elsewhere.)

graphs.py compiles and flake8 is clean. The maintained agent-facing guidance now lives in the penguin-pilot context pack.