[Afloat16]

Problem

EvalReranker always calls its scoring model, including when no query-document pairs remain. An empty retrieval run, zero reranking depth, or identical-ID filtering that removes all candidates reaches AbsReranker.compute_score([]) and raises an IndexError before evaluation can continue.

Changes

If the final pair list is empty, return an empty score mapping for each input query. Mixed batches that still contain candidates continue through the normal scoring path.

Validation

• Eight regression cases use a locally saved tiny BERT classifier and vocabulary through the actual BaseReranker and evaluation wrapper.
• Cover no queries, one/multiple empty candidate lists, all candidates removed by identical-ID filtering, zero reranking depth, and mixed empty/nonempty query groups.
• Exact final baseline: 6 failed, 2 passed. After the patch: 8 passed.
• New cases plus existing Trainer/import tests: 16 passed, 1 skipped. The skipped check requires Transformers 5.x.
• PyTorch 2.15.0+cpu, Transformers 4.57.1, FAISS CPU 1.16.0 and pytrec-eval-terrier 0.5.11; no remote checkpoint download or GPU run. Optional evaluation imports are skipped if their dependencies are absent; they were present for validation.
• This patch only handles an empty candidate set at the evaluation-wrapper boundary. Input candidate mutation and filtering depth remain separate behavior.
• Ruff and git diff --check pass.