D2 component 2: pool the pre-fusion recall arms into the deduped candidate set to be labeled for a query.

Input is the strategy_items map that escapes ranked_recall_items/6 — arm_name => ranked item list, each item shaped by memory_unit_to_recall_response/2 (string "id", "content", "context", and metadata["eval_id"] as the stable label key). We keep the top-K per arm, union across arms, dedup on the stable memory_ref (eval_id), and record EVERY arm's rank for each candidate (arm_ranks), plus the best rank and arm-name list.

arm_ranks is what makes the pool re-fusable offline: RRF scores a candidate from each arm's rank, so D2's coordinate-ascent can rescore a whole pool under a trial weighting without re-running recall. Pools recorded before this existed carry only best_rank and must be regenerated before they can be ascended against.

Candidates lacking a stable eval_id cannot be mapped to a reusable label, so they are dropped from the pool and returned under :dropped — the caller logs that as coverage loss rather than silently labeling against ephemeral ids. Pure: no DB, no recall call.