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/tests/test_joint.py

@@ -146,7 +146,7 @@ def _overall() -> dict[str, float]:
         "hit_at_10_wilson_lower": 0.45,
         "abstention_rate_answerable": 0.0,
         "abstention_rate_unanswerable": 0.0,
-        "wrong_holder_confidence_unanswerable": 0.2,
+        "top_hit_from_holder_unanswerable": 0.2,
     }
 
 

bench/tests/test_systems.py

@@ -162,3 +162,28 @@ def test_percentile_needs_timings_and_a_valid_fraction() -> None:
     with pytest.raises(SystemExit, match="percentile must be within"):
         runner.percentile(1.5)
     assert runner.percentile(0.5) >= 0.0
+
+
+def test_the_dense_model_is_loaded_once_per_process(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    """The query path used to rebuild the model for every query and measured that as retrieval latency."""
+    import sentence_transformers
+
+    from bench import systems
+
+    loaded: list[str] = []
+
+    class FakeModel:
+        def __init__(self, path: str, **_kwargs: object) -> None:
+            loaded.append(path)
+
+    monkeypatch.setattr(sentence_transformers, "SentenceTransformer", FakeModel)
+    systems._MODELS.clear()
+    model_dir = tmp_path / "dense"
+    model_dir.mkdir()
+    try:
+        assert systems._sentence_model(model_dir) is systems._sentence_model(model_dir)
+        assert loaded == [str(model_dir)]
+        with pytest.raises(SystemExit, match="model not downloaded"):
+            systems._sentence_model(tmp_path / "missing")
+    finally:
+        systems._MODELS.clear()