diff --git a/interactive_calibrator.py b/interactive_calibrator.py
index a6f6b85..4f28d3a 100755
--- a/interactive_calibrator.py
+++ b/interactive_calibrator.py
@@ -63,7 +63,10 @@ def camera_ptz() -> dict[str, int]:
 
 def remove_session_snapshots(paths: set[str]) -> None:
     for path in sorted(paths):
-        if not path.startswith("/smart-cam/data/scan/calibration/"):
+        if not path.startswith((
+            "/smart-cam/data/scan/manual/",
+            "/smart-cam/data/scan/calibration/",
+        )):
             continue
         subprocess.run(
             ["docker", "exec", "smart-cam", "rm", "-f", "--", path],
@@ -109,20 +112,31 @@ def choose_direction(axis: str, origin: dict[str, int], points: dict) -> int:
     return 1 if high - origin["tilt"] >= origin["tilt"] - low else -1
 
 
-def consistent_samples(reliable: list[tuple[int, float, float, dict]]) -> list:
-    """Return the largest group that estimates one mutually consistent FOV."""
+def consistent_samples(reliable: list[tuple[int, float, float, dict]],
+                       maximum_spread: float = 0.15) -> list:
+    """Return the largest group with a bounded end-to-end FOV spread.
+
+    The previous seed-radius clustering allowed one sample at -12% and
+    another at +12% into the same cluster.  That cluster could therefore have
+    nearly 24% end-to-end spread and was immediately rejected by the later
+    15% check.  Use sorted contiguous windows so membership and validation
+    apply the same definition of consistency.
+    """
     if not reliable:
         return []
-    clusters = [
-        [
-            sample for sample in reliable
-            if abs(sample[2] - seed[2]) / max(seed[2], 1.0) <= 0.12
-        ]
-        for seed in reliable
-    ]
-    return max(
-        clusters, key=lambda cluster: (len(cluster), -max(x[2] for x in cluster))
-    )
+    ordered = sorted(reliable, key=lambda sample: sample[2])
+    candidates = []
+    for begin in range(len(ordered)):
+        for end in range(begin, len(ordered)):
+            window = ordered[begin:end + 1]
+            _, spread = fov_statistics(window)
+            if spread <= maximum_spread:
+                candidates.append(window)
+    return max(candidates, key=lambda group: (
+        len(group),
+        -fov_statistics(group)[1],
+        sum(int(sample[3].get("inliers") or 0) for sample in group),
+    ))
 
 
 def fov_statistics(samples: list[tuple[int, float, float, dict]]) -> tuple[float, float]:
@@ -321,8 +335,17 @@ def estimate_axis(axis: str, origin: dict[str, int], points: dict,
             result.get("match_method") == "phase_correlation"
             and float(result.get("phase_response") or 0) >= core.PHASE_MIN_RESPONSE
         )
-        if (feature_reliable or phase_reliable) and 0.08 <= overlap <= 0.80:
-            reliable.append((delta, overlap, delta / (1.0 - overlap), result))
+        if (feature_reliable or phase_reliable) and 0.08 <= overlap <= 0.90:
+            # Hikvision absolute PTZ movement is quantized and can also have
+            # backlash.  At high optical zoom, using the requested delta
+            # instead of the SDK-reported actual movement creates a large,
+            # systematic FOV error.  Shortfall samples have already been
+            # rejected above, so the actual movement is authoritative here.
+            actual_delta = int(result.get("actual_axis_motion") or delta)
+            reliable.append((
+                actual_delta, overlap,
+                actual_delta / (1.0 - overlap), result,
+            ))
             if len(reliable) >= 4:
                 provisional = consistent_samples(reliable)
                 if len(provisional) >= 4:
@@ -372,7 +395,10 @@ def estimate_axis(axis: str, origin: dict[str, int], points: dict,
             for sample in consistent
         )
         allowed_spread = 0.08
-        if not strong_pair:
+        # Two independently measured FOVs that agree within 5% are already a
+        # strong geometric cross-check.  Do not reject such a pair merely
+        # because a natural scene supplied 74 rather than 100 ORB inliers.
+        if not strong_pair and spread > 0.05:
             core.move(origin)
             core.fail(
                 f"{axis}方向目前只有2个一致样本，但特征质量不足以使用双样本规则；"