[Afloat16]

Problem

The configured search or reranking depth is applied before excluded candidates are removed. Dense retrieval can return fewer eligible documents when a query ID also occurs in the corpus. BRIGHT retrieval has the same issue for its per-query exclusion lists, and reranking can discard a valid candidate at the depth boundary when an identical ID occupies an earlier slot.

Changes

• Retrieve one additional dense candidate when identical IDs are excluded, then cap the eligible output at the requested depth.
• For BRIGHT, retrieve enough candidates for the largest relevant exclusion list (plus an optional identical ID), bounded by the corpus size. Keep the existing N/A sentinel behavior.
• Remove identical IDs from the reranker candidate sequence before selecting its configured depth.

Validation

Added 30 tests using local tiny BERT models, the native embedder/reranker, and FAISS. Retrieval results are compared with a full NumPy similarity ranking after applying the same eligibility rules. Cases cover depths 1, 2, and larger than the corpus; multiple BRIGHT exclusions; queries with and without corpus IDs; and enabled/disabled identical-ID filtering.

• New tests against the unchanged upstream base: 14 failed, 16 passed.
• New tests with this change plus the existing Trainer/import checks: 38 passed, 1 skipped.
• Ruff on the new test module and git diff --check passed.

This change fixes candidate selection depth. It does not change embedding scores, reranking scores, or exclusion policy.