Tracking: path from methods freeze to first real audit

[esaran1]

Ordered work for collaborators. Do not skip ahead to paper numbers.

PR #1 (Enhance/aaai) already used GitHub number 1, so this tracker starts at #2.

1. Freeze and push the analysis codebase before any HPC or paper work #2 Freeze/push the analysis codebase
2. Preregister the paper estimand and seed-dependence assumptions #3 Preregister estimand and seed pairing
3. Define the actual recipes and what compute=1,2,4,8,16 means #4 Define recipes and compute accounting
4. Connect a real train/eval script to ASLA_REAL_TRAIN_EVAL #5 Connect the real trainer
5. Configure SLURM resources and cluster environment #6 Configure SLURM
6. Run the 45-job grid and record only measured BPB #7 Run 45 jobs; record measured BPB only
7. Run the first real audit: projection vs single-scale #8 First real audit vs single-scale
8. Parallel / later:
   • Empirical coverage study for bootstrap intervals #9 bootstrap coverage study (do not call ranker CIs conformal)
   • Decide paired-seed resampling if seeds are shared experimental units #10 paired-seed bootstrap if needed
   • Add Chinchilla uncertainty only if an identifiable N,D grid exists #11 Chinchilla uncertainty only if N,D exist
   • Specify the muP mechanism crossover predictor before implementing it #12 mechanism predictor spec (do not fake)
   • Build paper tables from frozen audit JSON, not from demo stdout #13 paper tables from JSON
   • Put uncertainty and budget roles on remaining figures #14 figure uncertainty
   • Pin a reproducible Python environment for Mac and HPC #15 environment pin
   • Independent reproduction and overclaim checklist before paper text #16 independent reproduction checklist

Current blocker: no real measured BPB in the repo. All 45 manifest rows are pending.