Document measured token efficiency

[robertkeus]

Summary

• replace qualitative token claims with pinned Honey and ESON benchmark results
• publish exact reproduction commands, revisions, statistical tests, quality results, and caveats
• add the graph-engineering sources behind Peter’s interpretation

What the evidence supports

• ESON: 28% fewer o200k tokens than compact JSON on five losslessly round-tripped handoff documents
• Honey: significant output and LOC reductions on the paired Opus 4.8 and GPT-5.5 suites
• total Honey cost: statistically inconclusive on both providers
• Peter: no end-to-end savings claim because there is no no-Honey/no-ESON control run

Verification

• recomputed full-opus48 and full-gpt55 from Green-PT/honey-for-devs@b39339e
• ran 29 ESON JS/Python tests plus bench:formats and bench:primer at Green-PT/honey-eson@d6809a1
• ./tests/install.sh
• ./tests/claude-smoke.sh
• git diff --check