fix(work_plane): project fit inliers from CV-frame pts_cv (not raw pc)

NuwaClient.pointcloud_roi's docstring claims OpenCV frame, but in
practice the raw payload arrives in Nuwa native frame (Z<0 forward).
nuwa_to_cv_frame is what actually flips it. Using `pts_cv` (post-flip
+ post-outlier-mask) makes the projection drop none of the inliers.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/presa_mano/api/work_plane_routes.py

@@ -115,8 +115,10 @@ def fit(req: FitReq) -> dict[str, Any]:
 
     # Project the inliers used for the fit back to RGB pixels so the GUI can
     # overlay them on the live image (verify visually which depth pixels were
-    # actually fed to fit_plane_svd). pc is in OpenCV camera frame already
-    # (per pointcloud_roi docstring); use Nuwa's RGB intrinsics.
+    # actually fed to fit_plane_svd). NOTE: pointcloud_roi's docstring claims
+    # CV frame, but in practice the raw `pc` arrives in Nuwa native frame
+    # (Z<0 forward). After nuwa_to_cv_frame + outlier mask, `pts_cv` is the
+    # canonical CV-frame inlier set, so use it.
     fit_points_px: list[list[int]] = []
     try:
         intr_full = nuwa.intrinsics()
@@ -127,9 +129,8 @@ def fit(req: FitReq) -> dict[str, Any]:
             H = int(rgb_intr.get("height", 480))
             rgb_intr = {**rgb_intr, "width": W, "height": H}
             K = K_from_intrinsics(rgb_intr)
-            inliers_cv = pc[inlier_mask, :3]   # original CV-frame coords
             with sw.step("project_inliers"):
-                fit_points_px = project_camera_to_pixels(inliers_cv, K, (W, H)).tolist()
+                fit_points_px = project_camera_to_pixels(pts_cv, K, (W, H)).tolist()
     except Exception as e:
         log.warning("could not project fit inliers to RGB pixels: %s", e)