CLAUDE.md optimization for lower AI agent token overhead

[RP2]

The CLAUDE.md is currently ~4,000 words (~5,000 tokens). Every time an AI agent loads the project, those tokens burn on descriptions that restate what's already in the code and the indexed knowledge graph.

I've been experimenting with a more compact "grug brain" style in my own opencode config with good results; inspired by projects like https://github.com/JuliusBrussee/caveman. The idea is to strip articles, helping verbs, and connector language without removing any constraints or code references. Agents parse the compressed form just as well, cutting the token count roughly in half.

Before putting time into a PR I wanted to check if there's any interest in this kind of change.

[marcinz606]

yes, I was meaning to also do that. Some time ago i added a bit up top

> **Keep this file current.** When a change alters something documented here — pipeline stage order, the feature pattern, the data model, build/packaging steps, or the dev commands — update the relevant section in the same change. After adding a pipeline stage, a feature, or a shader, re-check the "CPU pipeline" stage list and the "Adding a new feature" checklist.

but obviously it backfired with pace of new features and changes :) probably at this point majority of the CLAUDE.md should be cleaned up and this bit re-phrased to only update it with crucial pipeline changes

[marcinz606]

i cleaned it up significantly