Generate one rank per evaluation cohort

[nipunbatra]

What changed

• generate rank and a dedicated evaluation-cohort digest in the Python leaderboard builder
• keep sequence length, epochs, tuning, container, and runtime provenance in the full comparison record without letting them reset public rank
• render the generated rank directly in the site
• regenerate the checked-in JSON and CSV, yielding one cohort with ranks 1–17

Root cause

The browser grouped rows by the full comparison-protocol digest. That digest intentionally changes for model context length and runtime provenance, so DAE/BERT and Mean each started a new rank group despite using the same REDD/fridge evaluation cohort.

Verification

• uv run ruff check src tests
• uv run pytest — 129 passed, 1 skipped
• regenerated leaderboard JSON/CSV are byte-for-byte reproducible
• explicit regression coverage for differing sequence lengths and container revisions