Revert TTA: 3x inference timed out on HF Space (b6bc99a -> all zeros)

Set USE_TTA=False. The TTA strict variant locally showed +0.003 mean and
+0.029 q5, but the 3x inference cost exceeded HF Space's evaluation time
limit, returning 0.0 for every sample. Pipeline now runs the single-seed
fast path it used in 2df06c6 (the proven 0.4815 leaderboard config).

Also leaves USE_ENSEMBLE=False with infrastructure preserved: a fine-tune
attempt with heavy augmentation (jitter 0.01 + drop 0.1) produced a model
0.040 hss_mean WORSE than the original on local 50-sample A/B, so the
ensemble code in script.py + ensemble.py + tta.py is wired but disabled
pending a better-trained 2nd checkpoint.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

script.py

@@ -59,14 +59,25 @@ CONF_THRESH = 0.4
 MERGE_THRESH = 0.4
 SNAP_RADIUS = 0.5
 
-# Test-time augmentation: 3 priority-sample seeds + Hungarian matching with
-# strict 3-pass agreement (min_passes_for_keep=2). Local 100-sample A/B:
-# q5 0.062 -> 0.091 (+47%), mean +0.003. Costs 3x inference time but
-# strict filter dramatically improves precision on hard scenes.
-USE_TTA = True
+# Test-time augmentation: 3 priority-sample seeds + Hungarian matching.
+# Local 100-sample A/B was +0.003 mean / +0.029 q5 vs single-pass, BUT the
+# 3x inference cost (commit b6bc99a) timed out on HF Space and produced
+# all-zero scores. Disabled. Re-enable only with single-seed (2x not 3x)
+# or after profiling shows the HF Space can fit 3x within its time limit.
+USE_TTA = False
 TTA_SEEDS = (2718, 31415, 42)
 TTA_MIN_PASSES = 2
 
+# Multi-checkpoint ensemble: tried 2026-05-23 with checkpoint2.pt fine-tuned
+# under heavy aug (jitter 0.01 + drop 0.1) for 20k steps from checkpoint.pt.
+# Local 50-sample A/B (new model alone vs original alone): mean dropped by
+# 0.040 with 20 big losses vs 5 big wins -- the aug degraded model accuracy.
+# Ensemble of [original + new] therefore can't improve over original alone.
+# Keep flag disabled; need a more careful training run before re-enabling.
+USE_ENSEMBLE = False
+ENSEMBLE_MIN_PASSES = 1
+TTA_PLUS_ENSEMBLE_MIN_PASSES = 2
+
 
 def fuse_and_sample(sample, cfg, rng):
     """Run point fusion + priority sampling on a raw dataset sample.
@@ -380,6 +391,17 @@ if __name__ == "__main__":
     model = load_model(checkpoint_path, device)
     print(f"Model loaded: {sum(p.numel() for p in model.parameters()):,} params")
 
+    # Optional: load 2nd checkpoint for ensemble inference
+    ensemble_models = None
+    if USE_ENSEMBLE:
+        checkpoint2_path = SCRIPT_DIR / "checkpoint2.pt"
+        if checkpoint2_path.exists() and checkpoint2_path.stat().st_size > 1000:
+            model2 = load_model(checkpoint2_path, device)
+            ensemble_models = [model, model2]
+            print(f"Ensemble: loaded 2 models for cross-checkpoint averaging")
+        else:
+            print(f"USE_ENSEMBLE=True but checkpoint2.pt not present; running single-model")
+
     # Point fusion config
     cfg = FuserConfig()
     rng = np.random.RandomState(2718)
@@ -410,7 +432,44 @@ if __name__ == "__main__":
             pred_status = "ok"
             n_fused_pts = 0
 
-            if USE_TTA:
+            if ensemble_models is not None and USE_TTA:
+                # 2-model ensemble × multi-seed TTA: 2 * len(seeds) total passes.
+                # Strict cross-pass agreement filters spurious segments.
+                try:
+                    from ensemble import predict_sample_ensemble
+                    pred_v, pred_e = predict_sample_ensemble(
+                        sample, cfg, ensemble_models, device,
+                        seeds=TTA_SEEDS,
+                        min_passes_for_keep=TTA_PLUS_ENSEMBLE_MIN_PASSES,
+                    )
+                    if torch.cuda.is_available():
+                        torch.cuda.empty_cache()
+                except Exception as e:
+                    import traceback
+                    print(f"  Ensemble+TTA failed for {order_id}:\n{traceback.format_exc()}")
+                    pred_v, pred_e = empty_solution()
+                    pred_status = "ensemble_tta_failed"
+                    if torch.cuda.is_available():
+                        torch.cuda.empty_cache()
+            elif ensemble_models is not None:
+                # 2-model ensemble, single seed
+                try:
+                    from ensemble import predict_sample_ensemble
+                    pred_v, pred_e = predict_sample_ensemble(
+                        sample, cfg, ensemble_models, device,
+                        seeds=(2718,),
+                        min_passes_for_keep=ENSEMBLE_MIN_PASSES,
+                    )
+                    if torch.cuda.is_available():
+                        torch.cuda.empty_cache()
+                except Exception as e:
+                    import traceback
+                    print(f"  Ensemble failed for {order_id}:\n{traceback.format_exc()}")
+                    pred_v, pred_e = empty_solution()
+                    pred_status = "ensemble_failed"
+                    if torch.cuda.is_available():
+                        torch.cuda.empty_cache()
+            elif USE_TTA:
                 # Multi-seed TTA: fuse + predict 3 times, Hungarian-match segments
                 # across passes, drop those without min_passes agreement.
                 try: