Run the 45-job grid and record only measured BPB

[esaran1]

Depends on

• Freeze and push the analysis codebase before any HPC or paper work #2 methods freeze on GitHub
• Define the actual recipes and what compute=1,2,4,8,16 means #4 recipe/compute definitions
• Connect a real train/eval script to ASLA_REAL_TRAIN_EVAL #5 real trainer connected
• SLURM configured

Grid

Current data/run_manifest.csv:

• 3 interventions
• fit budgets 1,2,4
• reserved intermediate 8 (gate only; do not leak into projection fits)
• target 16
• 3 seeds
• 45 pending rows, 0 measured BPB

This HPC grid is not the synthetic demo ladder (1,2,4,8 / 16 / 64).

What to do

1. Execute the array with ASLA_EXECUTE=1 after dry-run inspection.
2. Collect:

python scripts/collect_results.py \
  --manifest data/run_manifest.csv \
  --results-dir results/hpc \
  --out data/run_manifest.csv

3. Finalize only when every row is completed with finite BPB:

python scripts/finalize_run_manifest.py \
  --manifest data/run_manifest.csv \
  --out-csv data/runs_template.csv \
  --out-parquet runs.parquet
python scripts/check_runs_coverage.py \
  --csv data/runs_template.csv \
  --target 16 \
  --fit-budgets 1 2 4 \
  --intermediate-budget 8
asla validate --runs runs.parquet

Acceptance

• All 45 rows status=completed with real bpb
• Coverage checker passes for target 16, fit 1 2 4, intermediate 8
• runs.parquet validates
• Failed jobs are rerun or explicitly excluded with a written reason; they are not imputed

Do not fabricate

• Fill BPB with placeholders, copies, or interpolations
• Bypass the finalizer
• Treat mock preflight bpb=1.234 as data