[log0u7]

Summary

• First end-to-end benchmark run published in docs/explanation/benchmark.md: config (3 models, local llama.cpp teacher on the 3080 for batch 2, judge_independence warn documented), data (23 CTF cases, 15 accepted), training (QLoRA Qwen3-1.7B, 6 pre-cutoff records), eval on the 9-record post-cutoff holdout.
• Numbers: base and student both 0.00 accuracy (prose output, no CtfVerdict JSON at the 6-record scale; the student degenerates into repetition loops). The honest reading: the pipeline is demonstrated end to end, the gain needs the 500-2000 record scale-up (ROADMAP phase 2).
• Four defects the run surfaced, all fixed and referenced (PR 24/25/26 + the warmup-vs-steps pitfall documented).
• README status row + ROADMAP benchmark item updated (ticked with the no-gain note), CHANGELOG entry.

Issue for this PR

N/A

Checklist

• Conventional Commits
• Docs only (benchmark.md, README, ROADMAP, CHANGELOG)