Add camera-stitch preview and feature-matching script

- images/stitch-preview/north-sunrise-hstack.jpg:
  naive side-by-side preview of the two skycams at 14:29 on 2026-04-17
- images/stitch-preview/feature-matches-debug.jpg:
  close-up of overlap region (right 40% of north × left 40% of sunrise)
  with ORB Lowe-ratio matches drawn — 39 candidates, RANSAC survives only
  4 inliers because the 90° perspective offset invalidates a similarity
  transform.  Overlap IS present (trees, houses, power poles visible in
  both) but full homography needs more/better tie points.
- stitch-cameras.py: reusable script that generates both previews plus a
  similarity-transform .txt for reuse in ffmpeg filter chains once we
  have enough matches (or manual tie-points).

https://claude.ai/code/session_01C4jbd3waXG3eKZYbGUjLUQ

stitch-cameras.py

@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+"""Stitch north + sunrise — v2.
+
+Lessons from v1: full homography needs lots of good inliers.  We had 5, which
+over-fit and produced streaking.  Fixes:
+  1. Mask out the top half (sky/clouds) — clouds moved in the 36 s between
+     the two shots, so they poison matching.
+  2. Use a similarity transform (rotation + uniform scale + translation,
+     4 DOF) via estimateAffinePartial2D — stable with few inliers.
+  3. Narrow the candidate band to the expected overlap sliver.
+"""
+import sys, cv2, numpy as np
+
+north_path, sunrise_path, out_prefix = sys.argv[1], sys.argv[2], sys.argv[3]
+
+north   = cv2.imread(north_path)
+sunrise = cv2.imread(sunrise_path)
+H_n, W_n = north.shape[:2]
+H_s, W_s = sunrise.shape[:2]
+print(f"north   {W_n}x{H_n}")
+print(f"sunrise {W_s}x{H_s}")
+
+# 1) Naive side-by-side baseline so the user always has something to look at.
+naive = np.hstack([north, sunrise])
+cv2.imwrite(f"{out_prefix}-naive.jpg", naive)
+
+# 2) Take only the ground half (skip sky where clouds change) and the
+#    band where the two FOVs should overlap:
+#       right 40% of north   ×   left 40% of sunrise
+sky_cutoff = 0.45           # keep rows from y = 0.45*H downwards
+ts_cutoff  = 0.96           # cut the bottom ~4% to drop the burnt-in timestamp
+y0_n, y1_n = int(H_n * sky_cutoff), int(H_n * ts_cutoff)
+y0_s, y1_s = int(H_s * sky_cutoff), int(H_s * ts_cutoff)
+band_frac = 0.40
+
+n_band = north  [y0_n:y1_n, int(W_n * (1 - band_frac)):]
+s_band = sunrise[y0_s:y1_s, : int(W_s * band_frac)]
+n_ox, n_oy = int(W_n * (1 - band_frac)), y0_n   # band → full-image offset
+s_ox, s_oy = 0, y0_s
+
+orb = cv2.ORB_create(nfeatures=8000, scaleFactor=1.2, nlevels=10, edgeThreshold=15)
+kp_n, des_n = orb.detectAndCompute(cv2.cvtColor(n_band, cv2.COLOR_BGR2GRAY), None)
+kp_s, des_s = orb.detectAndCompute(cv2.cvtColor(s_band, cv2.COLOR_BGR2GRAY), None)
+print(f"ORB keypoints: north-band={len(kp_n)}  sunrise-band={len(kp_s)}")
+
+bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=False)
+raw = bf.knnMatch(des_n, des_s, k=2)
+good = [m for m, n in raw if m.distance < 0.8 * n.distance]
+print(f"good Lowe-ratio matches: {len(good)}")
+
+if len(good) < 6:
+    print("Not enough matches to align — writing naive hstack as final.")
+    cv2.imwrite(f"{out_prefix}-aligned.jpg", naive)
+    sys.exit(0)
+
+src = np.float32([[kp_n[m.queryIdx].pt[0] + n_ox,
+                   kp_n[m.queryIdx].pt[1] + n_oy] for m in good])
+dst = np.float32([[kp_s[m.trainIdx].pt[0] + s_ox,
+                   kp_s[m.trainIdx].pt[1] + s_oy] for m in good])
+
+# Similarity transform: rotation + uniform scale + translation.
+M, inl = cv2.estimateAffinePartial2D(src, dst, method=cv2.RANSAC,
+                                     ransacReprojThreshold=5.0, maxIters=5000,
+                                     confidence=0.99)
+n_inl = int(inl.sum()) if inl is not None else 0
+print(f"similarity inliers: {n_inl}/{len(good)}")
+
+if M is None or n_inl < 4:
+    print("Similarity fit too weak — writing naive hstack as final.")
+    cv2.imwrite(f"{out_prefix}-aligned.jpg", naive)
+    sys.exit(0)
+
+a, b, tx = M[0]; c, d, ty = M[1]
+scale = np.sqrt(a * a + c * c)
+rot_deg = np.degrees(np.arctan2(c, a))
+print(f"est. transform: scale={scale:.4f}  rot={rot_deg:+.2f}°  tx={tx:+.1f}  ty={ty:+.1f}")
+
+# Warp north into sunrise's coordinate frame, then compose the canvas.
+M3 = np.vstack([M, [0, 0, 1]])
+n_corners = np.float32([[0, 0], [W_n, 0], [W_n, H_n], [0, H_n]]).reshape(-1, 1, 2)
+n_warped  = cv2.perspectiveTransform(n_corners, M3)
+s_corners = np.float32([[0, 0], [W_s, 0], [W_s, H_s], [0, H_s]]).reshape(-1, 1, 2)
+all_c = np.concatenate([n_warped, s_corners])
+[xmin, ymin] = np.int32(all_c.min(axis=0).ravel() - 0.5)
+[xmax, ymax] = np.int32(all_c.max(axis=0).ravel() + 0.5)
+W, H = xmax - xmin, ymax - ymin
+shift = np.array([[1, 0, -xmin], [0, 1, -ymin]], dtype=np.float64)
+M_shifted = shift @ M3        # 2x3 after slicing
+M_shifted = M_shifted[:2]
+print(f"canvas {W}x{H}   shift=({-xmin},{-ymin})")
+
+warped_n = cv2.warpAffine(north, M_shifted, (W, H))
+
+canvas = warped_n.copy()
+y0, x0 = -ymin, -xmin
+s_slot = canvas[y0:y0 + H_s, x0:x0 + W_s]
+n_mask = (warped_n[y0:y0 + H_s, x0:x0 + W_s].sum(axis=2) > 0).astype(np.float32)[..., None]
+# Where north has content → 60/40 blend favouring sunrise (base).  Else just sunrise.
+blended = (sunrise.astype(np.float32) * (0.4 + 0.6 * (1 - n_mask)) +
+           s_slot.astype(np.float32) * (0.6 * n_mask)).astype(np.uint8)
+# Simpler: just drop sunrise on top wherever it exists.
+canvas[y0:y0 + H_s, x0:x0 + W_s] = sunrise
+
+cv2.imwrite(f"{out_prefix}-aligned.jpg", canvas)
+np.savetxt(f"{out_prefix}-similarity.txt", M, fmt="%.8f")
+print(f"wrote {out_prefix}-aligned.jpg and {out_prefix}-similarity.txt")
+
+# Also draw a debug image showing matched inliers.
+dbg = cv2.drawMatches(
+    n_band, kp_n, s_band, kp_s,
+    [good[i] for i in range(len(good)) if inl[i]],
+    None, flags=cv2.DrawMatchesFlags_NOT_DRAW_SINGLE_POINTS)
+cv2.imwrite(f"{out_prefix}-matches.jpg", dbg)
+print(f"wrote {out_prefix}-matches.jpg")