noise-aware gate, frozen core of unseen frames, and a cycle script for the Spark

A paired bootstrap over the 149 unseen gold frames (2026-10-02) showed that a
difference between two models carries about +-0.023 of noise in F1, +-0.018 in
the mask score and +-0.10 in crystal F1, while the gate's tolerance was 0.005.

- tools/gate_stats.py: per-frame counts and a paired bootstrap (pure Python).
- train_run.py: when the baseline's gold predictions exist, the gate requires
  the candidate to beat it on F1 or the mask score in 90 % of the draws and
  not to lose the other in more than 90 %; the point rule is the fallback and
  is always reported (pass_point_rule).
- tools/compare_unseen.py replaces the Spark's ~/fair_gate_all.py: all models
  on the frames none trained on, on the frozen core (data/gold/unseen_core.json,
  149 frames) and on all unseen frames once the gold set grows, with the
  interval and win share against the active model.
- measure_weights.py keeps a version's gate verdict when it re-measures it.
- tools/spark_cycle.sh: gold growth, export, re-measure, train, gate, compare,
  push, with what cycles 3 and 4 taught built in.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

tests/test_gate.py

@@ -31,3 +31,49 @@ def test_gate_refuses_a_guarded_class_regression(tmp_path):
 
 def test_no_baseline_means_no_gate():
     assert train_run.gate(_report(0.8, 0.9, 1, 0), None) is None
+
+
+def _frames(per_frame_tp, fp=0, objects=2, iou=0.9):
+    """One class, ``objects`` gold objects per frame, ``tp`` of them found."""
+    return {f"f{i}": {0: (float(tp), float(fp), float(objects - tp), iou * tp)} for i, tp in enumerate(per_frame_tp)}
+
+
+def test_paired_bootstrap_separates_a_real_gain_from_a_tie():
+    from tools import gate_stats
+    base = _frames([1] * 60)
+    better = _frames([2] * 40 + [1] * 20)            # finds more on two thirds of the frames
+    same = _frames([1] * 60)
+    r = gate_stats.paired_bootstrap(better, base, draws=400)
+    assert r["frames"] == 60 and r["f1"]["win"] > 0.99 and r["f1"]["lo"] > 0 and r["mask_score"]["win"] > 0.99
+    t = gate_stats.paired_bootstrap(same, base, draws=400)
+    assert t["f1"]["delta"] == 0 and t["f1"]["win"] == 0.0     # never strictly better
+
+
+def test_paired_bootstrap_pairs_only_on_common_frames():
+    from tools import gate_stats
+    cand = _frames([2] * 10); base = _frames([1] * 6)
+    assert gate_stats.paired_bootstrap(cand, base, draws=50)["frames"] == 6
+    assert gate_stats.paired_bootstrap(cand, {}, draws=50) == {"frames": 0, "draws": 0}
+
+
+def test_gate_with_bootstrap_needs_a_clear_win(tmp_path):
+    base = tmp_path / "base.json"
+    base.write_text(json.dumps(_report(0.80, 0.90, 80, 20, {"loop_all": {"f1": 0.90}, "pin": {"f1": 0.90}})))
+    new = _report(0.806, 0.90, 81, 19, {"loop_all": {"f1": 0.90}, "pin": {"f1": 0.90}})
+    noisy = {"frames": 150, "draws": 2000, "f1": {"delta": 0.006, "lo": -0.02, "hi": 0.03, "win": 0.66},
+             "mask_score": {"delta": 0.004, "lo": -0.02, "hi": 0.03, "win": 0.60}}
+    g = train_run.gate(new, base, noisy)
+    assert g["pass_point_rule"] is True and g["pass"] is False and g["rule"] == "paired bootstrap"
+    clear = {"frames": 150, "draws": 2000, "f1": {"delta": 0.05, "lo": 0.02, "hi": 0.08, "win": 1.0},
+             "mask_score": {"delta": 0.04, "lo": 0.01, "hi": 0.07, "win": 0.99}}
+    assert train_run.gate(new, base, clear)["pass"] is True
+    lopsided = {"frames": 150, "draws": 2000, "f1": {"delta": 0.03, "lo": 0.0, "hi": 0.06, "win": 0.97},
+                "mask_score": {"delta": -0.04, "lo": -0.07, "hi": -0.01, "win": 0.01}}
+    assert train_run.gate(new, base, lopsided)["pass"] is False      # wins F1 but clearly loses the masks
+
+
+def test_gate_without_predictions_keeps_the_point_rule(tmp_path):
+    base = tmp_path / "base.json"
+    base.write_text(json.dumps(_report(0.69, 0.903, 60, 40)))
+    g = train_run.gate(_report(0.85, 0.883, 85, 15), base)
+    assert g["rule"] == "point" and g["pass"] is True and g["pass_point_rule"] is True