[Afloat16]

Problem

EvalReranker replaces each query's candidate mapping in the supplied search_results dictionary with a truncated top-k mapping. Reusing the same retrieval results for another reranker or depth silently loses candidates: a top-1 run followed by top-3 can only rerank the one candidate left by the first call.

Changes

Build the top-k candidate mappings locally without modifying the caller's retrieval results. Ranking, truncation, identical-ID filtering, and score computation keep their existing order and behavior.

Validation

• Eight native regression cases use a tiny BERT sequence classifier and local BERT vocabulary saved to disk, then run the actual BaseReranker inference path through EvalReranker.
• Check the original result values and nested mappings, top-1/top-2/top-4 selections with identical-ID filtering on/off, and comparisons at increasing/decreasing reranking depths.
• Exact final baseline: 7 failed, 1 passed. After the patch: 8 passed.
• New cases plus existing Trainer/import checks: 16 passed, 1 skipped. The skipped check requires Transformers 5.x.
• Tested with PyTorch 2.15.0+cpu, Transformers 4.57.1, FAISS CPU 1.16.0 and pytrec-eval-terrier 0.5.11. Tests skip cleanly when optional evaluation dependencies are absent; they were present for these runs. No remote model download or GPU execution was used.
• Ruff and git diff --check pass.