Phase 3: extrapolation to the Phase 1 box, progress, safety snapshots

Working copies live in logs/phase3/extrap/ on the run box (the scripts read and write there); this folder is the same code, kept in git. Run from the repository root with /venv/p2/bin/python -I.

extrap.py ROOT p2_metrics.json OUT: per run, accuracy, passes/block, expert loads/token, and tokens/s on the Phase 1 box (8x RTX PRO 6000 Server). Headline = same-box measurement:
- HumanEval, MBPP: today's judge (run_p3.sh judge arm = rd flags minus --full-redraft-every) decoded on every item on this box under the same 4-GPU load; old-box tok/s = today's judge measured there x rd's same-box speedup.
- Other five: today's judge on the first 12 items here (judge_sample); old-box tok/s = rd tok/s here (all items) x this box's slowdown on those 12 items. Checked against the full baselines: -2.4% / +0.4%; bootstrap 90% range with 12 items about +-4-8%, unbiased.
- Reference columns: the Phase 2 page's cost model, and a single 16-item calibration ratio.
calib.py: the 16-item setup-check calibration (same items, token-identical outputs on both boxes).
status.py: progress table (setup | benchmark | done/total | ETA IST | old-box tok/s).
build_artifact.py + page_template.html: the results page (Table 1 in the Phase 2 page's layout).
snapshot.py, snapshot_loop.sh: partial analysis + RESUME.md + HF/GitHub push every 15 min (5 min near the expected end of credit); same secret scan as backup_p3.py.
baseline_then_switch.sh, samples_then_rd.sh, switch_order.sh, order.txt: the baseline runs and the short-first benchmark order (humaneval mbpp aime gsm8k500 gpqa math500 ifeval541), applied between benchmarks.
p2_metrics.json: the Phase 2 page's per-benchmark metrics (stock, today's judge, CoDA, Qwen).