perf(reranker): add benchmark harness and passage clipping

- scripts/benchmark_reranker.py exercises the configured reranker
  with synthetic queries or live OpenSearch samples and prints
  p50/p95/p99 latency, mean latency, and pairs/sec throughput.
  Supports --warmup, --candidates, --passage-length, --source, and a
  --json-only mode for CI.
- app/indexing/reranker.py clips passages to 2048 characters before
  scoring so a runaway chunk cannot starve the cross-encoder beyond
  bge-reranker-v2-m3's training window.
- RUNBOOK.md gains a Reranker benchmark section with CPU/GPU SLO
  targets and a remediation ladder (lower top-K, raise batch size,
  switch device, disable reranker) when measured p95 exceeds budget.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

RUNBOOK.md

@@ -117,6 +117,40 @@ docker compose exec postgres psql -U legacyhub -d legacyhub -c \
    should not be rolled back casually. Restore from backup via the standard
    TeamHUB Suite backup runbook.
 
+## Reranker benchmark
+
+The reranker is the latency-defining stage of the hybrid search path. Run the
+benchmark on every hardware change (CPU vs GPU, instance type, batch size)
+before promoting the configuration.
+
+```bash
+# synthetic warmup + 32 queries x 40 candidates, ~700-char passages
+docker compose exec api python scripts/benchmark_reranker.py \
+  --queries 32 --candidates 40 --warmup 4
+
+# real corpus sample (after some documents are indexed)
+docker compose exec api python scripts/benchmark_reranker.py \
+  --source opensearch --query "ГОСТ 21.501-93" --candidates 40
+```
+
+Target SLOs (subject to revision once staging numbers land):
+
+| Metric              | CPU target | GPU target |
+|---------------------|-----------:|-----------:|
+| p95 latency / query |  < 700 ms  |  < 120 ms  |
+| Throughput          | > 60 pair/s | > 600 pair/s |
+
+If the measured p95 exceeds the budget, options in order of preference:
+
+1. Lower `RERANK_CANDIDATES` (default 40 — reducing to 20 roughly halves work).
+2. Increase `RERANKER_BATCH_SIZE` (memory permitting).
+3. Switch `RERANKER_DEVICE=cuda` and use a GPU-capable image.
+4. Disable reranker (`RERANKER_ENABLED=false`) and accept raw RRF order — the
+   API still returns useful results; the `reranked` field reports the truth.
+
+Passages are clipped to 2048 chars before being fed to the cross-encoder so a
+runaway chunk cannot starve the budget.
+
 ## Scaling notes (~70k PDFs)
 
 - Workers horizontally scale: `docker compose up -d --scale worker=8`.