feat(exp4): tune SAM-3 fusion on smoke evidence — dormant turf is LAWN

Measured on a dormant pilot tile: the SAM bare-dirt union overlaps the lawn
union on 42% of the tile with 1% unique -> dirt paints before lawn (lawn wins
overlaps, unique dirt survives). Driveway wins driveway/sidewalk overlaps;
driveway claims on the centerline buffer outside parcels reassign to road
(SAM swallows whole streets). Flag switches from raw grass/dirt overlap
(fires on every leaf-off tile) to unique-dirt >5%. Smoke-mode count message
fixed; --debug-masks added.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

scripts/exp4/autolabel_sam3.py

@@ -65,8 +65,12 @@ PROMPTS = {
     "road": ["paved road"],
     "vehicle": ["car"],
 }
-# Paint order: weakest first; later classes overwrite earlier ones where they overlap.
-PAINT_ORDER = ["lawn", "dirt", "tree", "driveway", "sidewalk", "road",
+# Paint order: weakest first; later classes overwrite earlier ones where they
+# overlap. Dirt paints BEFORE lawn: on leaf-off imagery SAM's "bare dirt" is a
+# near-duplicate of the dormant-lawn mask (measured 42% of a tile overlapping,
+# 1% unique), so lawn wins the overlap and dirt keeps only its unique regions.
+# Sidewalk before driveway for the same reason (driveway wins shared strips).
+PAINT_ORDER = ["dirt", "lawn", "tree", "sidewalk", "driveway", "road",
                "pool", "building", "vehicle"]
 
 
@@ -137,10 +141,14 @@ def fuse(sam: dict[str, np.ndarray], sup: dict[str, np.ndarray],
 
     masks = dict(sam)
     # Road needs the centerline prior or off-parcel evidence; on-parcel "road"
-    # pixels are usually driveway — reassign them.
-    road_ok = masks["road"] & (sup["road_prior"] | ~sup["parcel_any"])
-    masks["driveway"] = masks["driveway"] | (masks["road"] & sup["parcel_any"]
-                                             & ~sup["road_prior"])
+    # pixels are usually driveway — reassign them. Symmetrically, "driveway"
+    # claims on the centerline buffer outside any parcel are actually road
+    # (SAM's driveway concept swallows whole streets sometimes).
+    street_zone = sup["road_prior"] & ~sup["parcel_any"]
+    road_ok = (masks["road"] | (masks["driveway"] & street_zone)) \
+        & (sup["road_prior"] | ~sup["parcel_any"])
+    masks["driveway"] = (masks["driveway"] | (masks["road"] & sup["parcel_any"]
+                                              & ~sup["road_prior"])) & ~street_zone
     masks["road"] = road_ok
     # Building: require SAM + LiDAR agreement for auto-accept; either-alone is a
     # conflict zone left as ignore (the reviewer resolves it).
@@ -159,7 +167,9 @@ def fuse(sam: dict[str, np.ndarray], sup: dict[str, np.ndarray],
     conflicts = {
         "building_iou": round(float(b_agree.sum() / b_union), 3) if b_union else 1.0,
         "lawn_on_lidar_roof": round(float(((label == 1) & sup["lidar_building"]).sum() / lawn_px), 3),
-        "grass_dirt_overlap": round(float((sam["lawn"] & sam["dirt"]).sum() / total), 3),
+        # raw overlap is ~40% on dormant tiles (expected); only UNIQUE dirt is
+        # a signal worth human eyes when it's large
+        "dirt_unique": round(float((sam["dirt"] & ~sam["lawn"]).sum() / total), 3),
         "labeled_frac": round(float((label > 0).sum() / total), 3),
         "vehicle_frac": round(float((label == 9).sum() / total), 3),
     }
@@ -168,8 +178,8 @@ def fuse(sam: dict[str, np.ndarray], sup: dict[str, np.ndarray],
         reasons.append("building-mismatch")
     if conflicts["lawn_on_lidar_roof"] > 0.05:
         reasons.append("lawn-on-roof")
-    if conflicts["grass_dirt_overlap"] > 0.04:
-        reasons.append("grass-dirt-ambiguous")
+    if conflicts["dirt_unique"] > 0.05:
+        reasons.append("large-dirt-area")
     if conflicts["labeled_frac"] < 0.55:
         reasons.append("sparse-labels")
     conflicts["flag_reasons"] = reasons
@@ -195,6 +205,8 @@ def main() -> None:
     ap.add_argument("--data", required=True, help="crops folder from export_training_crops.py")
     ap.add_argument("--limit", type=int, default=0)
     ap.add_argument("--smoke", action="store_true", help="1 crop + raw API structure dump")
+    ap.add_argument("--debug-masks", action="store_true",
+                    help="save each class's raw SAM union to debug_masks/ (tuning aid)")
     ap.add_argument("--revision", default=None, help="pin facebook/sam3 revision")
     args = ap.parse_args()
     data = Path(args.data)
@@ -225,11 +237,11 @@ def main() -> None:
 
     ids = sorted(p.stem for p in (data / "images").glob("*.jpg"))
     todo = [i for i in ids if not (data / "labels" / f"{i}.png").exists()]
+    print(f"crops: {len(ids)} total, {len(ids) - len(todo)} already labeled, {len(todo)} to do")
     if args.smoke:
         todo = todo[:1]
     elif args.limit:
         todo = todo[: args.limit]
-    print(f"crops: {len(ids)} total, {len(ids) - len(todo)} already labeled, {len(todo)} to do")
 
     flags_path = data / "flags.csv"
     done = failed = flagged = 0
@@ -241,6 +253,11 @@ def main() -> None:
                 .resize((WORK_SIZE, WORK_SIZE), Image.LANCZOS)
             sup = load_supervision(data, crop_id, WORK_SIZE)
             sam = sam3_class_masks(model, processor, device, image, smoke=args.smoke)
+            if args.debug_masks:
+                (data / "debug_masks").mkdir(exist_ok=True)
+                for cls, m in sam.items():
+                    Image.fromarray((m * 255).astype(np.uint8)).save(
+                        data / "debug_masks" / f"{crop_id}_{cls}.png")
             label, conflicts = fuse(sam, sup, WORK_SIZE)
             save_outputs(data, crop_id, image, label, conflicts)
             if conflicts["flag_reasons"]: