Label the latest TabPFN-Rel versions on the leaderboard

[adrian-prior]

Generated by Claude Code

The README leaderboard and plot now label the current TabPFN-Rel versions tabpfn-rel-client-latest and tabpfn-rel-local-latest, using the aliases TabPFN-Rel 0.0.5 registers. The first local model is renamed from tabpfn-rel-local to tabpfn-rel-local-2026-08-15 to match the rename in TabPFN-Rel 0.0.5, and usage examples point at the -latest names. Ranks, Elo, and intervals are unchanged.

Details

Labels

• The dev group installs base tabpfn-rel>=0.0.5, which registers the TabPFN-Rel methods without the inference backends. CI gets the labels through --group dev, so no workflow change is needed.
• workflows/update_leaderboards.py fails when results.csv names a method that is not registered. Without that check, the labels would depend on which packages happen to be installed.
• Regenerated README.md and docs/leaderboards.png. In the generated tables only the TabPFN-Rel names change.

Rename of the first local model

• baseline_results/results.csv: the model value of the 63 tabpfn-rel-local rows becomes tabpfn-rel-local-2026-08-15. Checked with pandas: every other column matches main exactly.
• Descriptions of that specific run (tuning table, baseline sources, per-model appendix, README models table) use the dated name. The pip extra tabpfn-rel-local keeps its name.
• workflows/check_package_split.py checks the new registered name.

Usage examples
The README, docs/predictive-task.md, packages/relarena/README.md, the relbench_v1 docstring, and examples/olist_seller_churn.py now use tabpfn-rel-client-latest / tabpfn-rel-local-latest. Those examples now run the TabPFN 3.5 models instead of the 2026-08-15 ones.

Dependencies

• tabpfn-rel extras floor: 0.0.4 → 0.0.5. The lock moves only tabpfn-rel 0.0.4 → 0.0.5 and is written with uv 0.12.20. uv sync --locked passes with uv 0.12.20 and 0.11.7.
• The fixed fastdfs cutoff stays until 2026-10-01, when 1.2 clears the 3-day window.

Tests

• A new test checks that the generator refuses unregistered methods and writes nothing.
• The alias-move test keeps the installed aliases.
• Full suite on macOS: 471 passed, 6 skipped, 1 failed. The failure is an RT test that fails only on macOS (torch supports only the file_system sharing strategy there); this PR changes no RT code. pre-commit run --all-files passes.

🤖 Generated with Claude Code