Build paper tables from frozen audit JSON, not from demo stdout

[esaran1]

Why

asla demo is a narrative smoke test. asla report and scripts/run_paper_experiments.py exist, but paper numbers for the real grid should come from a frozen results/audit.json plus a deterministic table generator. Synthetic experiment scripts are not a substitute for the 45-job audit.

What to do

Add something like scripts/make_tables.py that reads audit JSON and writes:

• CSV and LaTeX for projection vs single-scale vs gate (if included); ensemble/race only if those rankers were actually evaluated on the same budget contract
• estimand name, n_boot, seed, input SHA-256, fit form, budget roles
• under-seeded cell list
• crossover table with fit is blind preserved as missing, not as a dash filled in later

Also generate a one-page methods box: resampling unit, noise-band definition, gate rule, and what claims are not licensed.

Acceptance

• Re-running the script on the same JSON is bit-identical
• Tables cannot be produced from --fast without labeling them as smoke tests
• None crossover budgets stay blank/None in LaTeX
• README documents the exact command used for the paper tables

Do not fabricate

• Copy numbers by hand from demo terminal output
• Leave Infinity or NaN in paper artifacts
• Treat scripts/run_paper_experiments.py --fast synthetic output as the real-grid result