Measure the counterpart honestly: warm models, vectorised search, scope split

Three defects made the external baseline numbers and labels wrong.

- bench/systems.py: _sentence_model rebuilt the model on every query, so dense arm
  latency was mostly model loading and the latency-vs-size curve was flat because a
  constant dominated it. Models are now cached per process.
- bench/counterpart.py: visible() and search() walked 404k Python objects per query.
  Per-item holder/source/time columns are built once and the query path is numpy;
  12.5 queries/s instead of 0.7.
- bench/metrics.py: the aggregate key was named wrong_holder_confidence_unanswerable
  while it counts top hits that ARE from the asking holder. BASELINE.md read that
  backwards. Renamed to top_hit_from_holder_unanswerable and the sentence rewritten.
- bench/run_external.py: the by-modality table read row['answerable'] but aggregate
  emits n_answerable, so the writer died after scoring finished. Each store is now
  scored twice, [holder scope] and [whole corpus], so the gain from the holder
  boundary is separable from the gain from ranking; progress lines every 500 queries.

Verification: pytest bench/tests -q -> 61 passed; ruff check bench/ -> all checks
passed; mypy (fresh cache) -> no issues in 25 source files.

bench/systems.py

@@ -168,13 +168,25 @@ class DenseIndex:
         return np.asarray(self._vectors @ vector[0], dtype=np.float64)
 
 
-def _sentence_model(model_dir: Path) -> Any:
-    """sentence-transformers ships no stubs, so the model is typed Any rather than a fake object."""
-    from sentence_transformers import SentenceTransformer
+_MODELS: dict[str, Any] = {}
 
-    if not model_dir.is_dir():
-        raise SystemExit(f"{model_dir}: model not downloaded, run python -m bench.fetch_models first")
-    return SentenceTransformer(str(model_dir), local_files_only=True)
+
+def _sentence_model(model_dir: Path) -> Any:
+    """Load a sentence-transformers model once per process.
+
+    The query path called this per query, and constructing the model from disk cost
+    more than the encode itself: dense arm p50 was 1.7s of which most was loading, and
+    the latency-vs-size curve came out flat because a constant dominated it.
+    sentence-transformers ships no stubs, so the model is typed Any rather than a fake object.
+    """
+    key = str(model_dir)
+    if key not in _MODELS:
+        from sentence_transformers import SentenceTransformer
+
+        if not model_dir.is_dir():
+            raise SystemExit(f"{model_dir}: model not downloaded, run python -m bench.fetch_models first")
+        _MODELS[key] = SentenceTransformer(key, local_files_only=True)
+    return _MODELS[key]
 
 
 def top_sources(scores: np.ndarray, atoms: Sequence[Atom], top_k: int) -> list[str]: