Bench autopilot tooling + 397B N=10 + GPU power analysis + unified cross-model

[Lightheartdevs]

Builds on the merged N=3 entry (#30). Adds a self-healing bench harness, extends 397B to N=10, adds GPU power analysis, and unifies the cross-model comparison on clean data.

Tooling (autonomous bench harness + /mmbt-bench skill)

• bench_autopilot.py — self-healing supervisor (endpoint restart, stuck-cell kill, idempotent resume, model presets, --publish, Pushover, heartbeat).
• bench_dashboard.py — color grid + --oneline/--watch/--html/--json/--trend/--flips/sparkline.
• bench_report.py — auto-generates the findings doc (scorecard + replicate-stability + redistribution + finish-reason audit).
• bench_power.py + gpu_power_logger.sh — NEW dual-GPU telemetry logger + analyzer.
• skill_draft/SKILL.md → installed as /mmbt-bench.

397B-A17B N=10 (240 cells, all done_signal)

• no-think 82/120, think 72/120 — same 7–8/12 band as Step-3.7-Flash, 27B-Q4, Coder-Next-Q4. Scale doesn't move the aggregate (now at N=10).
• Thinking net −10: helps p3_market (8→10) but craters p3_doc (9→2) and p3_pm (5→0). Per-task call, not a default.
• N=10 overturns small-N luck: p3_market no-think flips 1/3 (N=3) → 8/10 (auto-flagged in the stability table). The headline methodological result.
• Failure temperament tracks lineage, not size: 397B + 27B stall; Coder-Next + Flash run away.

Cross-model (Q4 refs, honest)

27B/Coder columns use the clean Q4/AWQ runs (microbench-2026-04-28). Fresh Q8/FP8 attempts excluded as serving failures (35B: 36/36 HTTP-400; 27B-Q8: 23/36 runaway) — documented, not faked.

GPU power — "are both GPUs ever near full power?" No (3,868 paired samples)

• Combined both-GPU: median 670W (56% of 1200W cap), max 985W (82%), 0 samples within 5% of full.
• GPU0 leads GPU1 (+27W) = pipeline-parallel alternation; decode 339W/GPU vs CPU-tool phases 125W/GPU.
• The "both at 600W" worst case doesn't occur for single-stream agentic inference under pipeline parallelism. (MiniMax-M2.7 TP=2 contrast pending in a follow-up commit.)

🤖 Generated with Claude Code

[Lightheartdevs]

Self-audit (against grade.json ground truth)

Audited the PR end-to-end. Findings:

• ✅ Scorecard accurate — recounted all 240 cells from grade.json: no-think 82/120, think 72/120, every per-task cell matches findings.md exactly.
• ✅ Power numbers reproduce — bench_power.py re-run matches power-analysis.md (max 985W=82% of cap, 0 samples within 5% of full; tiny mean drift is just more samples since commit).
• 🔴 Fixed staleness (the one real issue): README, QUALITATIVE.md, and manifest.json still carried N=3 numbers (23/36, 22/36, "net −1") while findings.md was N=10 — the entry was internally inconsistent. All three updated to N=10 (567ed15), with the sharpened finding (thinking net −10: market 8→10, doc 9→2, pm 5→0), plus the power + Q4-cross-model notes and the new file listings. No claims changed — only stale counts corrected.
• ✅ Tooling — all 5 scripts compile; reproduce commands match real flags. Removed the .v1.py pre-enrichment backups (26d8ce3) — git history preserves them; clutter in a public entry.
• ✅ No leaks — no run-logs, no broken-Q8 cells, no agent-pilot data in the diff.

Verdict: accurate and complete for everything through 397B N=10 + power + Q4 cross-model. MiniMax-M2.7 is deliberately not in this PR — it smoke-passed but the N=10 run 400-errored on every agentic cell (multi-turn serving-format issue); being debugged separately, will land as a follow-up so this PR isn't blocked.