Add TPP scaling-law example suite: data + goldens…

Details

A reproducible suite (mirroring pyedcr-examples/) that ships the paper's
published per-run loss CSVs and fits every case with the ORIGINAL fitting code
pinned at commit a63a008f, so nothing numeric is re-implemented in Python:
- fetch_data.py       sha256-pinned download of the 7 loss CSVs + the authors'
                      published per-seed parquet.
- common/paper_code.py locate/verify the pinned original code (env override or
                      on-demand clone); generation fails loudly off-commit.
- common/paper_numbers.py verbatim paper-quoted headline numbers, each tagged
                      with its .tex source.
- make_examples.py    orchestrates the original parsing / split / array-prep /
                      seed-aware fitters / coverage-step + bbox selection into
                      web/<dataset>.json (schema v1). Before trusting any golden
                      it re-fits every (law, mode, design) seed-0 combo and
                      requires BIT-EXACT agreement with the authors' published
                      results (100 combos verified); round-trips every case's
                      arrays back out of the JSON.
Shipped JSONs live at frontend/jekyll/assets/tpp-examples/ (served copy).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014c7dbCGbQxtWedKRSx4sJi

Files changed (24)