It feels wrong to dump identifiers to save tokens: now they're devoid of semantics, and can't be grep'ed or mapped to concepts. CPUs are good with numbers, but LLMs are good with words.