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>

app/indexing/reranker.py

@@ -50,10 +50,16 @@ class Reranker:
     def available(self) -> bool:
         return self._impl is not None and self._model is not None
 
+    # bge-reranker-v2-m3 is trained at 512 tokens; we truncate by chars so the
+    # reranker stays inside its budget even when callers forget to limit the
+    # candidate text length.
+    _MAX_PASSAGE_CHARS = 2048
+
     def score(self, query: str, passages: Sequence[str]) -> list[float]:
         if not self.available or not passages:
             return [0.0] * len(passages)
-        pairs = [(query, p) for p in passages]
+        clipped = [p[: self._MAX_PASSAGE_CHARS] for p in passages]
+        pairs = [(query, p) for p in clipped]
         if self._impl == "flagembedding":
             scores = self._model.compute_score(pairs, batch_size=self.batch_size, normalize=True)  # type: ignore[union-attr]
         else: