Controlled Pre-Reranking Evidence Coverage and the Downstream Value of Cross-Encoder Reranking in Multi-Hop RAG

Standard Retrieval-Augmented Generation (RAG) evaluations conflate upstream candidate starvation, boundary displacement, and generation failures. We isolate when and why cross-encoder reranking improves multi-hop QA in a paired study of $N = 1,700$ HotpotQA (dev_distractor) queries ($3,400$ generations) over |D| = 66,576 articles. Fixing the BM25 pool ($K = 20$), prompt, and generator (Google Gemini 2.5 Flash thinking model, logged under harness identifier gemini-3.8-flash; $T = 0.0$, max_tokens $= 1,024$), we compare BM25 top-5 (No-CE) against cross-encoder/ms-marco-MiniLM-L-6-v2 (CE; K = 20 → N = 5). Reranking improves Exact Match by = +8.12 pp (pHB = 3.52 × 10⁻²⁰) and token F₁ by Δ F₁ = +10.52 pp (pHB = 8.53 × 10⁻²⁹), with $86.1%$ concordant EM ties. Pre-reranking completeness gates efficacy: Complete Pools ($n = 1,111$) drive $97.1%$ of net EM wins (+12.06 pp, pHB = 5.30 × 10⁻²⁰), whereas Incomplete Pools ($n = 589$) yield +0.68 pp EM (pHB = 0.4240) and a borderline, family-dependent +1.23 pp F₁ (praw = 0.0340, pHB = 0.0340 at $m = 2$, pHB,Jt = 0.0680). Observational transitions show symmetric payoffs for PROMOTED ($n = 380$, +41.58 pp EM) vs. Bridge DEMOTED ($n = 34$, -41.18 pp EM), while REORDERED_ONLY ($n = 93$) yields 0.00 pp. Finally, telemetry reveals a $63:1$ thinking-to-visible token ratio ($782,227$ vs. $12,411$), requiring ample output budget to prevent silent truncation.