Sizes the rerank prefilter shortlist — how many fused candidates are handed to the (expensive) cross-encoder before the final selection.

Parity with Hindsight's rerank_limit = thinking_budget * 2 (memory_engine.py:3294): rerank the top 2× the resolved recall budget. An explicit recall_rerank_window override wins.

Deliberately NOT a function of the answer's max_tokens: Hindsight token-filters AFTER reranking (its step 5.5), so shrinking this shortlist by the answer cap would starve the cross-encoder the same way the pre-seed recall was starved before.