Re-measure arm latency on a warm-model sample instead of trusting the cache

rankings.jsonl stores the seconds each query took, so the report inherited whatever
the machine was doing at the time: the frozen baseline's dense latency included a
per-query model load, and the 50,000-source row moved from 1.315s to 0.809s between
two runs of identical code.

- bench/run.py: --remeasure-latency N re-times a stride sample with warm models and
  replaces each arm's latency section, recording the sample size and a note; recall
  stays a full-pass number. BASELINE.md prints that note under the table.
- bench/tests/test_joint.py: 62 tests total.

Verification: pytest bench/tests -q -> 62 passed; ruff check bench/ -> all checks
passed; mypy (fresh cache) -> no issues in 25 source files.

bench/tests/test_joint.py

@@ -8,6 +8,7 @@ from __future__ import annotations
 
 import json
 from pathlib import Path
+from typing import Any
 
 import pytest
 
@@ -17,6 +18,7 @@ from bench.run import (
     fingerprint,
     latency,
     load_rankings,
+    remeasure_latency,
     scale_arm_names,
     scale_probes,
     write_markdown,
@@ -200,3 +202,12 @@ def test_scale_arm_names_reads_only_latency_dicts() -> None:
     assert scale_arm_names(_report()["scale"]) == ["bm25_only"]
     with pytest.raises(SystemExit, match="no measured arm"):
         scale_arm_names([{"sources": 500, "note": "latency only"}])
+
+
+def test_remeasure_latency_replaces_cached_seconds_with_a_warm_sample() -> None:
+    arms: list[dict[str, Any]] = [{"arm": "bm25_only", "latency_seconds": {"p50": 99.0}}]
+    remeasure_latency(arms, JointRunner(["bm25_only"], ATOMS, systems.MODELS), _probe_set(8), 4)
+    measured = arms[0]["latency_seconds"]
+    assert measured["p50"] < 99.0
+    assert measured["queries"] == 4
+    assert "warm models" in measured["note"]