Summary

• split benchmark artifacts into reusable spec and per-run manifests
• add replay support, baseline-primary guidance, and model attribution to generation benchmarks
• extend benchmark reporting with platform, consumer, and model metadata

Testing

• node --test test/generation-session.test.mjs test/generation-benchmark.test.mjs

Notes

• Structural coverage is green locally.
• A live local-LLM qwen3.5-9b replay cohort is still pending as follow-up evidence.