setiastrosuitepro 1.6.7__py3-none-any.whl

setiastro/saspro/comet_stacking.py

@@ -0,0 +1,1442 @@
+# pro/comet_stacking.py
+from __future__ import annotations
+import os
+import sys
+import tempfile
+import subprocess
+import shutil
+import math
+import numpy as np
+import cv2
+from typing import List, Dict, Tuple, Optional
+from functools import lru_cache
+from astropy.io import fits
+from astropy.stats import sigma_clipped_stats
+import sep
+from setiastro.saspro.remove_stars import (
+    _get_setting_any,
+    _mtf_params_linked, _apply_mtf_linked_rgb, _invert_mtf_linked_rgb,
+    _resolve_darkstar_exe, _ensure_exec_bit, _purge_darkstar_io
+
+)
+from setiastro.saspro.legacy.image_manager import (load_image, save_image)
+from setiastro.saspro.imageops.stretch import stretch_color_image, stretch_mono_image
+
+def _blackpoint_nonzero(img_norm: np.ndarray, p: float = 0.1) -> float:
+    """Scalar blackpoint from non-zero pixels across all channels (linked).
+       p in [0..100]: small percentile to resist outliers; use 0 for strict min."""
+    x = img_norm
+    if x.ndim == 3 and x.shape[2] == 3:
+        nz = np.any(x > 0.0, axis=2)     # keep pixels where any channel has signal
+        vals = x[nz]                      # shape (N,3) → flatten to scalar pool
+    else:
+        vals = x[x > 0.0]
+    if vals.size == 0:
+        return float(np.min(x))           # fallback (all zeros?)
+    if p <= 0.0:
+        return float(np.min(vals))
+    return float(np.percentile(vals, p))
+
+def _float01_to_u16(x: np.ndarray) -> np.ndarray:
+    x = np.clip(np.asarray(x, dtype=np.float32), 0.0, 1.0)
+    return (x * 65535.0 + 0.5).astype(np.uint16, copy=False)
+
+def _u16_to_float01(x: np.ndarray) -> np.ndarray:
+    x = np.asarray(x)
+    dt = x.dtype
+
+    # Exact uint16 → normalize
+    if dt == np.uint16:
+        return (x.astype(np.float32) / 65535.0)
+
+    # TIFF/FITS readers sometimes return float32 0..65535
+    if dt in (np.float32, np.float64):
+        mx = float(np.nanmax(x)) if x.size else 0.0
+        if mx > 1.01:                       # looks like 0..65535
+            return (x.astype(np.float32) / 65535.0)
+        # already 0..1 (or very close) → just clip for safety
+        return np.clip(x.astype(np.float32), 0.0, 1.0)
+
+    # Be forgiving with 8-bit
+    if dt == np.uint8:
+        return (x.astype(np.float32) / 255.0)
+
+    # Fallback: assume 16-bit range
+    return (x.astype(np.float32) / 65535.0)
+
+# comet_stacking.py (or wherever this lives)
+
+def starnet_starless_pair_from_array(
+    src_rgb01,
+    settings,
+    *,
+    is_linear: bool,
+    debug_save_dir: str | None = None,
+    debug_tag: str | None = None,
+    core_mask: np.ndarray | None = None,   # <-- added (keyword-only)
+):
+    """
+    Standalone-like StarNet path using our imageops stretch:
+      - if linear:   stretch (per-channel) with 0.25  -> StarNet
+      - then:       pseudo-linear "unstretch" both orig & starless with 0.05
+    This avoids linked-MTF chroma issues and keeps both branches consistent.
+    """
+
+    exe = _get_setting_any(settings, ("starnet/exe_path", "paths/starnet"), "")
+    if not exe or not os.path.exists(exe):
+        raise RuntimeError("StarNet executable path is not configured.")
+    _ensure_exec_bit(exe)
+
+    # -------- normalize & shape: float32 [0..1], keep note if mono ----------
+    x = np.asarray(src_rgb01, dtype=np.float32)
+    was_mono = (x.ndim == 2) or (x.ndim == 3 and x.shape[2] == 1)
+    
+    # DELAY expansion: work with 'x' (mono or rgb) directly where possible
+    x_input = x
+    if x_input.ndim == 3 and x_input.shape[2] == 1:
+        x_input = x_input[..., 0] # collapse to 2D for processing if needed, or keep 2D
+
+    # For StarNet save, we need 3 channels usually, but check if we can save mono?
+    # Actually StarNet usually expects RGB Tiff. So we might need to expand just for saving.
+    # But let's avoid `x3 = np.repeat` globally.
+    
+    # Optimization: Create x3 ON DEMAND or virtually using broadcasting only when needed.
+    # But `save_image` might handle mono TIFs. If StarNet accepts Mono TIF, we save huge RAM.
+    # Standard StarNet typically wants RGB. We will enable "is_mono" flag in `save_image` if it is mono,
+    # but StarNet is finicky. Let's stick to RGB for StarNet input but avoid `np.repeat` for the WHOLE array 
+    # if we can just broadcast or slice. 
+    # Actually, `stretch_color_image` handles broadcasting? No.
+    # Let's simple optimize:
+    
+    if is_linear:
+         # stretch; if mono use mono stretch
+         if was_mono:
+             if x.ndim == 3: x = x[..., 0]
+             pre = stretch_mono_image(x, 0.25, False, False)
+             # expand ONLY for save
+             pre_to_save = np.dstack([pre, pre, pre])
+         else:
+             pre = stretch_color_image(x, 0.25, False, False, False)
+             pre_to_save = pre
+    else:
+        pre = x # floating point 0..1
+        if was_mono:
+            if pre.ndim == 3: pre = pre[..., 0]
+            pre_to_save = np.dstack([pre, pre, pre])
+        else:
+            pre_to_save = pre
+
+    # -------- StarNet I/O (write float->16b TIFF; read back float) ----------
+    starnet_dir = os.path.dirname(exe) or os.getcwd()
+    in_path  = os.path.join(starnet_dir, "imagetoremovestars.tif")
+    out_path = os.path.join(starnet_dir, "starless.tif")
+
+    save_image(pre_to_save, in_path, original_format="tif", bit_depth="16-bit",
+                   original_header=None, is_mono=False, image_meta=None, file_meta=None)
+
+    exe_name = os.path.basename(exe).lower()
+    if os.name == "nt" or sys.platform.startswith(("linux","linux2")):
+        cmd = [exe, in_path, out_path, "256"]
+    else:
+        cmd = [exe, "--input", in_path, "--output", out_path] if "starnet2" in exe_name else [exe, in_path, out_path]
+
+    rc = subprocess.call(cmd, cwd=starnet_dir)
+    if rc != 0 or not os.path.exists(out_path):
+        try: os.remove(in_path)
+        except Exception as e:
+            import logging
+            logging.debug(f"Exception suppressed: {type(e).__name__}: {e}")
+        raise RuntimeError(f"StarNet failed (rc={rc}).")
+
+    starless_pre, _, _, _ = load_image(out_path)
+    try:
+        os.remove(in_path); os.remove(out_path)
+    except Exception:
+        pass
+
+    # Don't expand starless_pre yet if we don't need to.
+    starless_pre = starless_pre.astype(np.float32, copy=False)
+    if was_mono and starless_pre.ndim == 3:
+        # StarNet output is usually RGB even for mono input. Convert back to mono?
+        # Or just use one channel.
+        starless_pre = starless_pre[..., 0]
+    
+    # Maintain `pre` as the stretched input (mono or rgb)
+    
+    # ---- mask-protect in the SAME (stretched) domain as pre/starless_pre ----
+    if core_mask is not None:
+        m = np.clip(core_mask.astype(np.float32), 0.0, 1.0)
+        # broadcast mask
+        if not was_mono:
+             if m.ndim == 2: m = m[..., None]
+        
+        protected_stretched = starless_pre * (1.0 - m) + pre * m
+    else:
+        protected_stretched = starless_pre
+    
+    # Return to 3-channel ONLY if requested by the caller's context? 
+    # The signature `starnet_starless_pair_from_array` implies it might return what it got.
+    # The original returned `protected_unstretch`.
+    pass # logic flow continues below...
+
+    # -------- “unstretch” → shared pseudo-linear space (once, after blend) ----------
+    if is_linear:
+        # choose stretcher based on channels
+        if was_mono:
+             # ensure 2d
+             if protected_stretched.ndim == 3 and protected_stretched.shape[2] == 1:
+                 protected_stretched = protected_stretched[..., 0]
+             elif protected_stretched.ndim == 3:
+                 # collapse rgb to mono if needed? likely StarNet gave RGB.
+                 # Keep RGB if StarNet created color artifacts we want to keep? 
+                 # Usually for mono data we want to kill color.
+                 protected_stretched = protected_stretched.mean(axis=2)
+             
+             protected_unstretch = stretch_mono_image(
+                protected_stretched, 0.05, False, False
+             )
+             # Expand finally for return constraint? 
+             # The older function returned RGB-like. 
+             # Let's expand here at the VERY END.
+             protected_unstretch = np.dstack([protected_unstretch]*3)
+        else:
+             protected_unstretch = stretch_color_image(
+                protected_stretched, 0.05, False, False, False
+             )
+    else:
+        protected_unstretch = protected_stretched
+        if was_mono and protected_unstretch.ndim == 2:
+            protected_unstretch = np.dstack([protected_unstretch]*3)
+
+    return np.clip(protected_unstretch, 0.0, 1.0), np.clip(protected_unstretch, 0.0, 1.0)
+
+
+
+def darkstar_starless_from_array(src_rgb01: np.ndarray, settings, **_ignored) -> np.ndarray:
+    """
+    Headless CosmicClarity DarkStar run for a single RGB frame.
+    Returns starless RGB in [0..1]. Uses CC’s input/output folders.
+    """
+    # normalize channels
+    img = src_rgb01.astype(np.float32, copy=False)
+    # Delay expansion: if it's 2D/Mono, send it as-is if DarkStar supports it, 
+    # but DarkStar expects 3-channel TIF usually. 
+    # We'll just expand for the save call, not "in place" if possible.
+    # Actually DarkStar runner saves `img` directly.
+    # So we'll expand just for that save to avoid holding 2 copies in memory.
+    if img.ndim == 2: 
+        img_to_save = np.stack([img]*3, axis=-1)
+    elif img.ndim == 3 and img.shape[2] == 1: 
+        img_to_save = np.repeat(img, 3, axis=2)
+    else:
+        img_to_save = img
+
+    # resolve exe and base folder
+    exe, base = _resolve_darkstar_exe(type("Dummy", (), {"settings": settings})())
+    if not exe or not base:
+        raise RuntimeError("Cosmic Clarity DarkStar executable path is not set.")
+
+    _ensure_exec_bit(exe)
+
+    input_dir  = os.path.join(base, "input")
+    output_dir = os.path.join(base, "output")
+    os.makedirs(input_dir, exist_ok=True)
+    os.makedirs(output_dir, exist_ok=True)
+
+    # purge any prior files (safe; scoped to imagetoremovestars*)
+    _purge_darkstar_io(base, prefix="imagetoremovestars", clear_input=True, clear_output=True)
+
+    in_path  = os.path.join(input_dir, "imagetoremovestars.tif")
+    out_path = os.path.join(output_dir, "imagetoremovestars_starless.tif")
+
+    # save input as float32 TIFF
+    # save input as float32 TIFF
+    save_image(img_to_save, in_path, original_format="tif", bit_depth="32-bit floating point",
+                   original_header=None, is_mono=False, image_meta=None, file_meta=None)
+
+    # build command (SASv2 parity): default unscreen, show extracted stars off, stride 512
+    cmd = [exe, "--star_removal_mode", "unscreen", "--chunk_size", "512"]
+
+    rc = subprocess.call(cmd, cwd=output_dir)
+    if rc != 0 or not os.path.exists(out_path):
+        try: os.remove(in_path)
+        except Exception as e:
+            import logging
+            logging.debug(f"Exception suppressed: {type(e).__name__}: {e}")
+        raise RuntimeError(f"DarkStar failed (rc={rc}).")
+
+    starless, _, _, _ = load_image(out_path)
+    # cleanup
+    try:
+        os.remove(in_path)
+        os.remove(out_path)
+        _purge_darkstar_io(base, prefix="imagetoremovestars", clear_input=True, clear_output=True)
+    except Exception:
+        pass
+
+    if starless is None:
+        raise RuntimeError("DarkStar produced no output.")
+
+    # Delayed expansion
+    if starless.ndim == 2: 
+        starless = np.stack([starless]*3, axis=-1)
+    elif starless.ndim == 3 and starless.shape[2] == 1: 
+        starless = np.repeat(starless, 3, axis=2)
+        
+    return np.clip(starless.astype(np.float32, copy=False), 0.0, 1.0)
+
+# ---------- small helpers ----------
+def _inv_affine_2x3(M: np.ndarray) -> np.ndarray:
+    """Invert a 2x3 affine matrix [[a,b,tx],[c,d,ty]] → [[a',b',tx'],[c',d',ty']]."""
+    A = np.asarray(M, dtype=np.float64).reshape(2,3)
+    a,b,tx = A[0]; c,d,ty = A[1]
+    det = a*d - b*c
+    if abs(det) < 1e-12:
+        raise ValueError("Affine matrix not invertible")
+    inv = np.array([[ d, -b, 0.0],
+                    [-c,  a, 0.0]], dtype=np.float64) / det
+    # new translation = - inv * t
+    inv[:,2] = -inv[:,:2] @ np.array([tx, ty], dtype=np.float64)
+    return inv.astype(np.float32)
+
+def _to_luma(img: np.ndarray) -> np.ndarray:
+    if img.ndim == 2: return img.astype(np.float32, copy=False)
+    if img.ndim == 3 and img.shape[-1] == 3:
+        try:
+            return cv2.cvtColor(img, cv2.COLOR_RGB2GRAY).astype(np.float32, copy=False)
+        except Exception:
+            pass
+        r,g,b = img[...,0], img[...,1], img[...,2]
+        return (0.2126*r + 0.7152*g + 0.0722*b).astype(np.float32, copy=False)
+    if img.ndim == 3 and img.shape[-1] == 1:
+        return img[...,0].astype(np.float32, copy=False)
+    return img.astype(np.float32, copy=False)
+
+def _robust_centroid(img: np.ndarray, seed_xy: Optional[Tuple[float,float]]=None, r=40) -> Optional[Tuple[float,float]]:
+    """Find a compact bright blob near seed using SEP; fallback to image max."""
+    L = _to_luma(img)
+    H,W = L.shape
+    if seed_xy:
+        x0,y0 = int(round(seed_xy[0])), int(round(seed_xy[1]))
+        x1,x2 = max(0,x0-r), min(W, x0+r+1)
+        y1,y2 = max(0,y0-r), min(H, y0+r+1)
+        roi = L[y1:y2, x1:x2]
+        if roi.size >= 16:
+            bkg = np.median(roi)
+            try:
+                sep.set_extract_pixstack(int(1e6))
+                objs, seg = sep.extract(roi - bkg, thresh=2.0*np.std(roi), minarea=8, filter_type='matched')
+                if len(objs):
+                    # pick highest peak
+                    k = int(np.argmax([o['peak'] for o in objs]))
+                    cx = float(objs[k]['x']) + x1
+                    cy = float(objs[k]['y']) + y1
+                    return (cx, cy)
+            except Exception:
+                pass
+    # fallback: global maximum
+    j = int(np.argmax(L))
+    cy, cx = divmod(j, W)
+    return (float(cx), float(cy))
+
+def _star_suppress(L: np.ndarray) -> np.ndarray:
+    """Down-weight stellar pinpoints so big fuzzy cores win."""
+    small = cv2.GaussianBlur(L, (0, 0), 1.6).astype(np.float32)
+    thr = np.percentile(small, 99.7)
+    mask = small > thr              # very bright, compact stuff
+    out = L.astype(np.float32, copy=True)
+    out[mask] *= 0.35               # damp stars; keep coma
+    return out
+
+def _log_big_blob(L: np.ndarray, sigmas: list[float]) -> tuple[float, float, float]:
+    """
+    Pick the strongest bright blob across multiple scales using LoG-like response.
+    Returns (cx, cy, sigma_used).
+    """
+    H, W = L.shape
+    best_val, best_xy, best_s = -1e9, (W*0.5, H*0.5), sigmas[0]
+    for s in sigmas:
+        g  = cv2.GaussianBlur(L, (0, 0), s)
+        lap = cv2.Laplacian(g, cv2.CV_32F, ksize=3)
+        resp = (-lap) * (s * s)     # scale-normalized: favor larger bright blobs
+        hi = np.percentile(resp, 99.95)
+        resp = np.clip(resp, -1e9, hi)
+        j = int(np.argmax(resp))
+        cy, cx = divmod(j, W)
+        v = resp[cy, cx]
+        if v > best_val:
+            best_val, best_xy, best_s = float(v), (float(cx), float(cy)), float(s)
+    return best_xy[0], best_xy[1], best_s
+
+
+# --- NEW helpers ---
+def _luma_gauss(img: np.ndarray, sigma: float=3.0) -> np.ndarray:
+    L = _to_luma(img)
+    return cv2.GaussianBlur(L, (0,0), sigmaX=sigma, sigmaY=sigma).astype(np.float32, copy=False)
+
+def _crop_bounds(cx, cy, half, W, H):
+    x1 = max(0, int(round(cx - half)))
+    y1 = max(0, int(round(cy - half)))
+    x2 = min(W, int(round(cx + half)))
+    y2 = min(H, int(round(cy + half)))
+    return x1, y1, x2, y2
+
+def _norm_patch(p: np.ndarray) -> np.ndarray:
+    m = np.median(p)
+    s = np.std(p)
+    if s < 1e-6: s = 1e-6
+    return ((p - m) / s).astype(np.float32, copy=False)
+
+def _minmax_time_key(fp: str) -> float:
+    # Try FITS DATE-OBS; fallback to file mtime. Lower is earlier.
+    try:
+        hdr = fits.getheader(fp, 0)
+        t = hdr.get("DATE-OBS") or hdr.get("DATE")
+        if t:
+            # robust parse: YYYY-MM-DDThh:mm:ss[.sss][Z]
+            from datetime import datetime
+            for fmt in ("%Y-%m-%dT%H:%M:%S.%f", "%Y-%m-%dT%H:%M:%S", "%Y/%m/%d %H:%M:%S"):
+                try:
+                    return datetime.strptime(t.replace("Z",""), fmt).timestamp()
+                except Exception:
+                    pass
+    except Exception:
+        pass
+    try:
+        return os.path.getmtime(fp)
+    except Exception:
+        return 0.0
+
+def _predict(prev_xy: Tuple[float,float], prev2_xy: Optional[Tuple[float,float]]) -> Tuple[float,float]:
+    if prev2_xy is None: 
+        return prev_xy
+    vx = prev_xy[0] - prev2_xy[0]
+    vy = prev_xy[1] - prev2_xy[1]
+    return (prev_xy[0] + vx, prev_xy[1] + vy)
+
+# --- NEW per-frame star masks (optional, safer than warping) ---
+def build_star_masks_per_frame(file_list: List[str], sigma: float=3.5, dilate_px: int=2, status_cb=None) -> Dict[str, np.ndarray]:
+    log = status_cb or (lambda *_: None)
+    masks = {}
+    k = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (2*dilate_px+1, 2*dilate_px+1)) if dilate_px>0 else None
+    for fp in file_list:
+        img, _, _, _ = load_image(fp)
+        if img is None: 
+            log(f"  ⚠️ mask: failed to load {os.path.basename(fp)}"); 
+            continue
+        L = _to_luma(img)
+        bkg, _, std = sigma_clipped_stats(L, sigma=3.0, maxiters=5)
+        m = (L > (bkg + sigma*std)).astype(np.uint8)
+        if k is not None:
+            m = cv2.dilate(m, k)
+        masks[fp] = (m > 0)
+        log(f"  ◦ star mask made for {os.path.basename(fp)}")
+    return masks
+
+@lru_cache(maxsize=32)
+def _directional_gaussian_kernel(long_px: int, sig_long: float,
+                                 sig_cross: float, angle_deg: float) -> np.ndarray:
+    """
+    Anisotropic Gaussian (elongated) rotated to `angle_deg`.
+    long_px controls kernel size along the tail axis.
+    Results are cached for reuse.
+    """
+    long_px = max(21, int(long_px) | 1)
+    half = long_px // 2
+    yy, xx = np.mgrid[-half:half+1, -half:half+1].astype(np.float32)
+    # rotate coords
+    th = np.deg2rad(angle_deg)
+    xr =  np.cos(th)*xx + np.sin(th)*yy     # along-tail
+    yr = -np.sin(th)*xx + np.cos(th)*yy     # cross-tail
+    g = np.exp(-0.5*( (xr/sig_long)**2 + (yr/sig_cross)**2 ))
+    g /= g.sum()
+    return g.astype(np.float32)
+
+def _anisotropic_feather(mask_bin: np.ndarray,
+                         angle_deg: float,
+                         feather_long: float,
+                         feather_cross: float) -> np.ndarray:
+    """
+    Feather with different falloff along vs. across tail by convolving
+    the binary mask with an elongated Gaussian oriented at angle_deg.
+    """
+    k = _directional_gaussian_kernel(
+        long_px=int(max(31, 6*max(feather_long, feather_cross))),
+        sig_long=float(max(1.0, feather_long/2.5)),
+        sig_cross=float(max(1.0, feather_cross/2.5)),
+        angle_deg=angle_deg
+    )
+    soft = cv2.filter2D(mask_bin.astype(np.float32), -1, k, borderType=cv2.BORDER_REPLICATE)
+    return np.clip(soft, 0.0, 1.0).astype(np.float32)
+
+def _tail_response(L: np.ndarray, angle_deg: float,
+                   bg_sigma: float = 30.0,
+                   hp_sigma: float = 2.0,
+                   long_px: int = 181,
+                   sig_long: float = 40.0,
+                   sig_cross: float = 3.0) -> np.ndarray:
+    """
+    Build a smooth tail-likelihood map: high-pass -> directional blur
+    (elongated Gaussian) -> normalize to [0,1].
+    """
+    # remove large-scale gradient, keep positive high-pass
+    low = cv2.GaussianBlur(L, (0,0), bg_sigma)
+    hp  = L - low
+    hp  = cv2.GaussianBlur(hp, (0,0), hp_sigma)
+    hp[hp < 0] = 0.0
+    k  = _directional_gaussian_kernel(long_px, sig_long, sig_cross, angle_deg)
+    resp = cv2.filter2D(hp, -1, k, borderType=cv2.BORDER_REFLECT)
+    # robust scale
+    p1, p99 = np.percentile(resp, (1.0, 99.7))
+    if p99 <= p1: 
+        return np.zeros_like(resp, np.float32)
+    return np.clip((resp - p1) / (p99 - p1), 0.0, 1.0).astype(np.float32)
+
+# At top of file (or near other imports)
+from setiastro.saspro.imageops.stretch import stretch_color_image, stretch_mono_image
+
+def _ensure_rgb_float01(x: np.ndarray) -> np.ndarray:
+    x = np.asarray(x)
+    if x.ndim == 2:
+        x = np.stack([x]*3, axis=-1)
+    elif x.ndim == 3 and x.shape[2] == 1:
+        x = np.repeat(x, 3, axis=2)
+    x = x.astype(np.float32, copy=False)
+    x = np.nan_to_num(x, nan=0.0, posinf=0.0, neginf=0.0)
+    return np.clip(x, 0.0, 1.0)
+
+def _ensure_mono_float01(x: np.ndarray) -> np.ndarray:
+    x = np.asarray(x)
+    if x.ndim == 3 and x.shape[2] == 3:
+        x = x.mean(axis=2)
+    elif x.ndim == 3 and x.shape[2] == 1:
+        x = x[..., 0]
+    x = x.astype(np.float32, copy=False)
+    x = np.nan_to_num(x, nan=0.0, posinf=0.0, neginf=0.0)
+    return np.clip(x, 0.0, 1.0)
+
+def blend_screen_stretched(
+    comet_only: np.ndarray,
+    stars_only: np.ndarray,
+    *,
+    stretch_pct: float = 0.05,  # use 5% like you requested
+    mix: float = 1.0,           # 0..1, scales the comet contribution in the screen
+) -> np.ndarray:
+    """
+    Display-stretch both inputs with your imageops stretch, then screen blend:
+      screen(A,B) = A + B - A*B
+    We apply 'mix' only to the comet term: out = screen(mix*A, B).
+
+    Returns float32 [0..1], RGB if any input is RGB, otherwise mono.
+    """
+
+    A = np.asarray(comet_only)
+    B = np.asarray(stars_only)
+
+    is_rgb = (A.ndim == 3 and A.shape[-1] == 3) or (B.ndim == 3 and B.shape[-1] == 3)
+
+    # 1) normalize/rgb-mono handling
+    if is_rgb:
+        A = _ensure_rgb_float01(A)
+        B = _ensure_rgb_float01(B)
+        # 2) stretch each with your display stretch (no links, no extra ops)
+        A_s = stretch_color_image(A, stretch_pct, False, False, False).astype(np.float32, copy=False)
+        B_s = stretch_color_image(B, stretch_pct, False, False, False).astype(np.float32, copy=False)
+    else:
+        A = _ensure_mono_float01(A)
+        B = _ensure_mono_float01(B)
+        A_s = stretch_mono_image(A, stretch_pct, False, False).astype(np.float32, copy=False)
+        B_s = stretch_mono_image(B, stretch_pct, False, False).astype(np.float32, copy=False)
+
+    # 3) screen blend with comet mix:
+    #    screen(mix*A, B) = B + mix*A - (mix*A)*B
+    mix = float(np.clip(mix, 0.0, 1.0))
+    out_s = B_s + mix*A_s - (mix*A_s)*B_s
+
+    return np.clip(out_s, 0.0, 1.0).astype(np.float32, copy=False)
+
+
+# --- REPLACE measure_comet_positions with this version ---
+def measure_comet_positions(
+    file_list: List[str],
+    seeds: Optional[Dict[str, Tuple[float,float]]] = None,
+    status_cb=None,
+    *,
+    tpl_half: int = 28,
+    blur_sigma: float = 3.5,
+    max_step_px: float = 45.0,
+    min_search_px: float = 16.0,
+    max_search_px: float = 80.0,
+    score_floor: float = 0.35,
+    gamma_pow: float = 0.6,
+    refine_r: int = 12,
+    adapt_tpl_alpha: float = 0.12
+) -> Dict[str, Tuple[float,float]]:
+    """
+    Track the comet by template matching on blurred luma.
+    Frames are processed in temporal order (DATE-OBS; fallback mtime).
+
+    Now with a SECOND PASS local refinement that mirrors the Comet preview “Auto” button.
+    """
+    log = status_cb or (lambda *_: None)
+
+    # -------- PASS 1: existing template-matching pipeline (unchanged) --------
+    ordered = sorted(list(file_list), key=_minmax_time_key)
+    out: Dict[str, Tuple[float,float]] = {}
+    prev_xy: Optional[Tuple[float,float]] = None
+    prev2_xy: Optional[Tuple[float,float]] = None
+    tpl: Optional[np.ndarray] = None
+    tpl_hw = int(tpl_half)
+
+    # Seed selection logic (unchanged)
+    seed_idx = 0
+    if seeds:
+        for i, f in enumerate(ordered):
+            if f in seeds:
+                seed_idx = i
+                break
+
+    for i, fp in enumerate(ordered):
+        img, hdr, _, _ = load_image(fp)
+        if img is None:
+            log(f"⚠️ measure: failed to load {fp}")
+            continue
+
+        # blurred luma + gamma for detection
+        L = _luma_gauss(img, sigma=blur_sigma)       # float32
+        G = _gamma_stretch(L, gamma=gamma_pow)       # [0..1]
+        H, W = G.shape
+
+        if tpl is None:
+            # choose seed
+            if seeds and fp in seeds:
+                cx, cy = seeds[fp]
+            elif seeds:
+                for f in ordered:
+                    if f in seeds: cx, cy = seeds[f]; break
+            else:
+                j = int(np.argmax(G)); cy, cx = divmod(j, W)
+
+            # keep user/global seed as the first output; refine subpixel on original luma (gamma’d)
+            L0g = _gamma_stretch(_to_luma(img), gamma=gamma_pow)
+            cx, cy = _refine_centroid(L0g, float(cx), float(cy), r=refine_r)
+
+            x1,y1,x2,y2 = _crop_bounds(cx, cy, tpl_half, W, H)
+            tpl = _norm_patch(G[y1:y2, x1:x2])
+            prev_xy = (float(cx), float(cy))
+            out[fp] = prev_xy
+            log(f"  ◦ seed @ {os.path.basename(fp)} → ({prev_xy[0]:.2f},{prev_xy[1]:.2f}) [template {tpl.shape[1]}×{tpl.shape[0]}]")
+            continue
+
+        # prediction & adaptive search window
+        guess = _predict(prev_xy, prev2_xy)
+        if prev2_xy is None:
+            sr = max(min_search_px, 0.5*max_step_px)
+        else:
+            mv = math.hypot(prev_xy[0]-prev2_xy[0], prev_xy[1]-prev2_xy[1])
+            sr = np.clip(1.5*mv, min_search_px, max_search_px)
+
+        # ensure search ≥ template
+        min_half_needed = 0.5 * max(tpl.shape[1], tpl.shape[0]) + 1.0
+        sr = max(sr, min_half_needed)
+
+        # crop and match
+        x1, y1, x2, y2 = _bounds_with_min_size(guess[0], guess[1], sr, W, H,
+                                               min_w=tpl.shape[1], min_h=tpl.shape[0])
+        search = _norm_patch(G[y1:y2, x1:x2])
+        res = cv2.matchTemplate(search, tpl, cv2.TM_CCOEFF_NORMED)
+        _, score, _, loc = cv2.minMaxLoc(res)
+        px = x1 + loc[0] + tpl.shape[1]*0.5
+        py = y1 + loc[1] + tpl.shape[0]*0.5
+
+        step = math.hypot(px - prev_xy[0], py - prev_xy[1])
+        ok = (score >= score_floor) and (step <= max_step_px)
+
+        if not ok:
+            # one wider search
+            x1b, y1b, x2b, y2b = _bounds_with_min_size(guess[0], guess[1], max_search_px, W, H,
+                                                       min_w=tpl.shape[1], min_h=tpl.shape[0])
+            search2 = _norm_patch(G[y1b:y2b, x1b:x2b])
+            res2 = cv2.matchTemplate(search2, tpl, cv2.TM_CCOEFF_NORMED)
+            _, score2, _, loc2 = cv2.minMaxLoc(res2)
+            px2 = x1b + loc2[0] + tpl.shape[1]*0.5
+            py2 = y1b + loc2[1] + tpl.shape[0]*0.5
+            step2 = math.hypot(px2 - prev_xy[0], py2 - prev_xy[1])
+            if (score2 > score) and (step2 <= max_step_px*1.2):
+                px, py, score, step = px2, py2, score2, step2
+                ok = (score >= 0.30)
+
+        if not ok:
+            px, py = _predict(prev_xy, prev2_xy)
+            px = float(np.clip(px, 0, W-1)); py = float(np.clip(py, 0, H-1))
+            log(f"  ◦ {os.path.basename(fp)} fallback → ({px:.2f},{py:.2f})")
+        else:
+            # subpixel refine on original luma (gamma’d)
+            L0 = _to_luma(img)
+            L0g = _gamma_stretch(L0, gamma=gamma_pow)
+            px, py = _refine_centroid(L0g, px, py, r=refine_r)
+            log(f"  ◦ {os.path.basename(fp)} match={score:.3f} step={step:.1f}px → ({px:.2f},{py:.2f})")
+
+            # gentle template adaptation
+            x1t, y1t, x2t, y2t = _crop_bounds(px, py, tpl_half, W, H)
+            new_tpl = _norm_patch(G[y1t:y2t, x1t:x2t])
+            if new_tpl.shape == tpl.shape:
+                tpl = (1.0 - adapt_tpl_alpha) * tpl + adapt_tpl_alpha * new_tpl
+
+        out[fp] = (px, py)
+        prev2_xy, prev_xy = prev_xy, (px, py)
+
+    # light smoothing (unchanged)
+    if len(out) >= 5:
+        ordered_xy = [out[f] for f in ordered]
+        xs = np.array([p[0] for p in ordered_xy], dtype=np.float64)
+        ys = np.array([p[1] for p in ordered_xy], dtype=np.float64)
+        def _smooth(v):
+            s = v.copy()
+            for k in range(2, len(v)-2):
+                s[k] = (-3*v[k-2] + 12*v[k-1] + 17*v[k] + 12*v[k+1] - 3*v[k+2]) / 35.0
+            return s
+        xs, ys = _smooth(xs), _smooth(ys)
+        for f, x, y in zip(ordered, xs, ys):
+            out[f] = (float(x), float(y))
+
+    # -------- PASS 2: local “Auto” refinement around first-pass XY --------
+    # Mirrors the dialog’s Auto: star-suppress → multi-scale LoG peak → gamma → subpixel refine
+    hint = max(4.0, blur_sigma)                    # reuse blur as the size hint
+    sigmas = [0.6*hint, 0.9*hint, 1.3*hint, 1.8*hint, 2.4*hint]
+    local_half = int(max(24, 3.0*hint))            # tight local window
+
+    for fp in ordered:
+        if fp not in out:
+            continue
+        img, _, _, _ = load_image(fp)
+        if img is None:
+            continue
+        Lfull = _to_luma(img).astype(np.float32)
+        cx0, cy0 = out[fp]
+        x1, y1, x2, y2 = _crop_bounds(cx0, cy0, local_half, Lfull.shape[1], Lfull.shape[0])
+
+        # star-suppressed local area + LoG peak
+        Ls = _star_suppress(Lfull[y1:y2, x1:x2])
+        cx, cy, used = _log_big_blob(Ls, sigmas)
+        cx += x1; cy += y1
+
+        # gamma + subpixel refine on the full-luma gamma space
+        gL = _gamma_stretch(Lfull, gamma=gamma_pow)
+        cx, cy = _refine_centroid(gL, float(cx), float(cy), r=max(refine_r, int(used)))
+
+        out[fp] = (float(cx), float(cy))
+
+    # light re-smoothing (keeps trajectories silky)
+    if len(out) >= 5:
+        ordered_xy = [out[f] for f in ordered]
+        xs = np.array([p[0] for p in ordered_xy], dtype=np.float64)
+        ys = np.array([p[1] for p in ordered_xy], dtype=np.float64)
+        def _smooth(v):
+            s = v.copy()
+            for k in range(2, len(v)-2):
+                s[k] = (-3*v[k-2] + 12*v[k-1] + 17*v[k] + 12*v[k+1] - 3*v[k+2]) / 35.0
+            return s
+        xs, ys = _smooth(xs), _smooth(ys)
+        for f, x, y in zip(ordered, xs, ys):
+            out[f] = (float(x), float(y))
+
+    return out
+
+
+def _bounds_with_min_size(cx, cy, half, W, H, min_w, min_h):
+    # Start from requested half-size
+    half = max(half, 1.0)
+    # First pass crop
+    x1 = int(round(cx - half)); y1 = int(round(cy - half))
+    x2 = int(round(cx + half)); y2 = int(round(cy + half))
+    # Clamp to image
+    x1 = max(0, x1); y1 = max(0, y1)
+    x2 = min(W, x2); y2 = min(H, y2)
+
+    # Ensure minimum width/height by expanding/shift-in if needed
+    cur_w = x2 - x1; cur_h = y2 - y1
+    need_w = max(0, int(min_w - cur_w))
+    need_h = max(0, int(min_h - cur_h))
+
+    # Expand symmetrically where possible; otherwise shift inward from edges
+    if need_w > 0:
+        x1 = max(0, x1 - need_w // 2)
+        x2 = min(W, x2 + (need_w - (x1 > 0 and (need_w // 2))))
+        # If still short, push entirely to one side
+        if (x2 - x1) < min_w:
+            if x1 == 0: x2 = min(W, min_w)
+            if x2 == W: x1 = max(0, W - min_w)
+
+    if need_h > 0:
+        y1 = max(0, y1 - need_h // 2)
+        y2 = min(H, y2 + (need_h - (y1 > 0 and (need_h // 2))))
+        if (y2 - y1) < min_h:
+            if y1 == 0: y2 = min(H, min_h)
+            if y2 == H: y1 = max(0, H - min_h)
+
+    # Final clamp/sanity
+    x1 = max(0, min(x1, W))
+    x2 = max(0, min(x2, W))
+    y1 = max(0, min(y1, H))
+    y2 = max(0, min(y2, H))
+    return x1, y1, x2, y2
+
+
+def build_star_masks_from_ref(ref_path: str,
+                              ref_star_thresh_sigma: float,
+                              inv_transforms: Dict[str, np.ndarray],
+                              dilate_px: int = 2,
+                              status_cb=None) -> Dict[str, np.ndarray]:
+    """Detect stars in ref, then warp mask back to each frame using inverse affine."""
+    log = status_cb or (lambda *_: None)
+    ref_img, hdr, _, _ = load_image(ref_path)
+    L = _to_luma(ref_img)
+    bkg, _, std = sigma_clipped_stats(L, sigma=3.0, maxiters=5)
+    thresh = bkg + ref_star_thresh_sigma * std
+    mask_ref = (L > thresh).astype(np.uint8)
+    if dilate_px > 0:
+        k = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (2*dilate_px+1, 2*dilate_px+1))
+        mask_ref = cv2.dilate(mask_ref, k)
+
+    H, W = L.shape
+    masks = {}
+    for f, Minv in inv_transforms.items():
+        m = cv2.warpAffine(mask_ref, Minv, (W, H),
+                           flags=cv2.INTER_NEAREST, borderMode=cv2.BORDER_CONSTANT, borderValue=0)
+        masks[f] = m.astype(bool, copy=False)
+        log(f"  ◦ star mask warped for {os.path.basename(f)}")
+    return masks
+
+def _shift_to_comet(img: np.ndarray, xy: Tuple[float,float], ref_xy: Tuple[float,float]) -> np.ndarray:
+    """Translate image so comet xy → ref_xy (subpixel)."""
+    dx = ref_xy[0] - xy[0]
+    dy = ref_xy[1] - xy[1]
+    M = np.array([[1.0, 0.0, dx], [0.0, 1.0, dy]], dtype=np.float32)
+    H, W = img.shape[:2]
+    interp = cv2.INTER_LANCZOS4
+    
+    # Vectorized warp for both 2D (mono) and 3D (RGB)
+    return cv2.warpAffine(img, M, (W, H), flags=interp, borderMode=cv2.BORDER_REFLECT)
+
+def stack_comet_aligned(file_list: List[str],
+                        comet_xy: Dict[str, Tuple[float,float]],
+                        star_masks: Optional[Dict[str, np.ndarray]] = None,
+                        reducer: str = "biweight",
+                        status_cb=None,
+                        *,
+                        settings=None,
+                        enable_star_removal: bool = False,
+                        star_removal_tool: str = "StarNet",
+                        core_r_px: float = 22.0,
+                        core_soft_px: float = 6.0,
+                        frames_are_linear: bool = True) -> np.ndarray:
+    """
+    If enable_star_removal=True, each comet-aligned frame has stars removed
+    with the chosen tool and nucleus protected by a soft circular mask.
+    """
+    log = status_cb or (lambda *_: None)
+    ref_xy = comet_xy[file_list[0]]
+
+    accum = []
+    core_mask_cache = None
+
+    for fp in file_list:
+        img, hdr, _, _ = load_image(fp)
+        if img is None: continue
+
+        shifted = _shift_to_comet(img, comet_xy[fp], ref_xy).astype(np.float32)
+
+        if enable_star_removal:
+            h, w = shifted.shape[:2]
+            if core_mask_cache is None:
+                # mask centered at ref_xy after shifting (all frames share this center now)
+                core_mask_cache = _protect_core_mask(h, w, ref_xy[0], ref_xy[1], core_r_px, core_soft_px)
+            shifted = _starless_frame_for_comet(
+                shifted, star_removal_tool, settings,
+                is_linear=frames_are_linear, core_mask=core_mask_cache
+            )
+            # after removal, star_masks are usually unnecessary; ignore them
+        else:
+            # keep your existing optional masks if not removing stars
+            if star_masks and fp in star_masks:
+                m = star_masks[fp]
+                if shifted.ndim == 2:
+                    shifted[m] = np.nan
+                else:
+                    for c in range(shifted.shape[-1]): shifted[...,c][m] = np.nan
+
+        accum.append(shifted)
+
+    if not accum:
+        raise RuntimeError("No valid frames for comet stacking")
+
+    stack = np.stack(accum, axis=0)
+
+    # same reducer as before
+    if reducer == "median":
+        out = np.nanmedian(stack, axis=0)
+    else:
+        med = np.nanmedian(stack, axis=0)
+        mad = np.nanmedian(np.abs(stack - med), axis=0) + 1e-8
+        k = 3.0
+        lo, hi = med - k*1.4826*mad, med + k*1.4826*mad
+        clipped = np.clip(stack, lo, hi)
+        out = np.nanmean(clipped, axis=0)
+    return out.astype(np.float32, copy=False)
+
+def make_comet_mask(comet_only: np.ndarray, feather_px: int=24) -> np.ndarray:
+    L = _to_luma(comet_only)
+    bkg, _, std = sigma_clipped_stats(L, sigma=3.0, maxiters=5)
+    m = (L > (bkg + 1.2*std)).astype(np.uint8)
+    # binary close + distance feather
+    k = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (7,7))
+    m = cv2.morphologyEx(m, cv2.MORPH_CLOSE, k, iterations=1)
+    # feather via distance transform
+    inv = 1 - m
+    dist = cv2.distanceTransform(inv, cv2.DIST_L2, 5)
+    mask = np.clip(1.0 - dist / max(1, feather_px), 0.0, 1.0)
+    return mask.astype(np.float32)
+
+
+
+# --- estimate global streak angle from comet motion (deg) ---
+def _estimate_streak_angle(comet_xy: dict[str, tuple[float,float]]) -> float:
+    if not comet_xy or len(comet_xy) < 2:
+        return 0.0
+    # order by time-ish from filename sort (good enough here)
+    ks = sorted(comet_xy.keys())
+    x0, y0 = comet_xy[ks[0]]
+    x1, y1 = comet_xy[ks[-1]]
+    # stars streak opposite comet motion; angle in image coords
+    ang = math.degrees(math.atan2(y0 - y1, x0 - x1))  # y down
+    return ang
+
+def _line_kernel(length: int, angle_deg: float) -> np.ndarray:
+    """Thin line (1px) dilated to ~3px width; rotated to angle."""
+    length = max(3, int(length))
+    w = 3
+    k = np.zeros((length, length), np.uint8)
+    cv2.line(k, (0, length//2), (length-1, length//2), 1, 1)
+    M = cv2.getRotationMatrix2D((length/2-0.5, length/2-0.5), angle_deg, 1.0)
+    rsz = cv2.warpAffine(k*255, M, (length, length), flags=cv2.INTER_NEAREST)
+    if w > 1:
+        rsz = cv2.dilate(rsz, cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(w,w)))
+    return (rsz > 0).astype(np.uint8)
+
+def _streak_mask_directional(comet_only: np.ndarray,
+                             angle_deg: float,
+                             hp_sigma: float = 2.0,
+                             bg_sigma: float = 15.0,
+                             th_sigma: float = 3.0,
+                             line_len: int = 19,
+                             grow_px: int = 2) -> np.ndarray:
+    """
+    Detect elongated bright streaks roughly along 'angle_deg'.
+    Returns boolean mask (H,W) where True = streak.
+    """
+    L = _to_luma(comet_only).astype(np.float32)
+    # high-pass: remove large-scale coma/tail
+    low = cv2.GaussianBlur(L, (0,0), bg_sigma)
+    hp  = cv2.GaussianBlur(L - low, (0,0), hp_sigma)
+
+    # robust threshold via MAD
+    med = np.median(hp)
+    mad = np.median(np.abs(hp - med)) + 1e-6
+    z = (hp - med) / (1.4826 * mad)
+    m0 = (z > th_sigma).astype(np.uint8)
+
+    # directional opening to keep long, aligned features; suppress compact bits
+    kline = _line_kernel(line_len, angle_deg)
+    opened = cv2.morphologyEx(m0, cv2.MORPH_OPEN, kline)
+
+    # small cleanups
+    if grow_px > 0:
+        k = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (2*grow_px+1, 2*grow_px+1))
+        opened = cv2.dilate(opened, k)
+    opened = cv2.morphologyEx(opened, cv2.MORPH_CLOSE,
+                              cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5,5)))
+    return opened.astype(bool)
+
+def _comet_mask_smart(comet_only: np.ndarray,
+                      feather_px: int,
+                      exclude_mask: np.ndarray | None = None,
+                      sigma_k: float = 1.2) -> np.ndarray:
+    """
+    Stronger comet mask: threshold broad coma/tail, remove star streaks,
+    then feather edges by distance.
+    """
+    L = _to_luma(comet_only).astype(np.float32)
+    bkg, _, std = sigma_clipped_stats(L, sigma=3.0, maxiters=5)
+    base = (L > (bkg + sigma_k * std)).astype(np.uint8)
+
+    # clean & expand a bit so tail isn’t holey
+    base = cv2.morphologyEx(base, cv2.MORPH_CLOSE,
+                            cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(9,9)), iterations=1)
+    base = cv2.dilate(base, cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(5,5)), iterations=1)
+
+    if exclude_mask is not None:
+        base[exclude_mask] = 0
+
+    # feather
+    inv = 1 - base
+    dist = cv2.distanceTransform(inv, cv2.DIST_L2, 5)
+    mask = np.clip(1.0 - dist / max(1, float(feather_px)), 0.0, 1.0)
+    return mask.astype(np.float32)
+
+def make_comet_mask_anisotropic(comet_only: np.ndarray,
+                                angle_deg: float,
+                                *,
+                                core_k: float = 1.2,
+                                tail_boost: float = 0.7,
+                                exclude_streaks: np.ndarray | None = None,
+                                feather_long: float = 90.0,
+                                feather_cross: float = 18.0) -> np.ndarray:
+    """
+    Tail-aware comet matte:
+      1) core/inner coma via sigma threshold,
+      2) add a directional tail likelihood,
+      3) remove star streaks,
+      4) anisotropic feather along tail.
+    """
+    L = _to_luma(comet_only).astype(np.float32)
+    bkg, _, std = sigma_clipped_stats(L, sigma=3.0, maxiters=5)
+    core = (L > (bkg + core_k*std)).astype(np.uint8)
+
+    # grow core a touch so it’s not holey around nucleus
+    core = cv2.morphologyEx(core, cv2.MORPH_CLOSE, cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(7,7)))
+    core = cv2.dilate(core, cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(3,3)), 1)
+
+    # directional tail map ∈ [0,1]; boost then clamp
+    tail = _tail_response(L, angle_deg=angle_deg)
+    tail = np.clip(tail * float(tail_boost), 0.0, 1.0)
+
+    # combine: binarize core strongly, add soft tail
+    m0 = np.clip(core.astype(np.float32) * 1.0 + tail * (1.0 - core.astype(np.float32)), 0.0, 1.0)
+
+    # remove linear star streaks if provided
+    if exclude_streaks is not None:
+        m0[exclude_streaks] = 0.0
+
+    # hard floor to keep nucleus fully in
+    m_bin = (m0 > 0.15).astype(np.uint8)
+
+    # anisotropic feather (stretches along tail, tight across)
+    matte = _anisotropic_feather(m_bin, angle_deg=angle_deg,
+                                 feather_long=feather_long,
+                                 feather_cross=feather_cross)
+    return np.clip(matte, 0.0, 1.0).astype(np.float32)
+
+def blend_comet_stars(
+    comet_only: np.ndarray,
+    stars_only: np.ndarray,
+    feather_px: int = 24,   # kept for compatibility; now used as cross-feather
+    mix: float = 1.0,
+    *,
+    comet_xy: dict[str, tuple[float,float]] | None = None
+) -> np.ndarray:
+    """
+    Tail-aware blend. Uses directional matte instead of radial blob.
+    `feather_px` controls *cross-tail* softness; along-tail uses a longer value automatically.
+    """
+    A = np.asarray(comet_only, dtype=np.float32)
+    B = np.asarray(stars_only, dtype=np.float32)
+
+    # channel harmonization
+    ch = 3 if ((A.ndim==3 and A.shape[-1]==3) or (B.ndim==3 and B.shape[-1]==3)) else 1
+    if ch == 3:
+        if A.ndim == 2: A = np.repeat(A[...,None], 3, axis=2)
+        if B.ndim == 2: B = np.repeat(B[...,None], 3, axis=2)
+    else:
+        if A.ndim == 3 and A.shape[-1] == 1: A = A[...,0]
+        if B.ndim == 3 and B.shape[-1] == 1: B = B[...,0]
+
+    angle = _estimate_streak_angle(comet_xy) if comet_xy else 0.0
+    # streak mask (same as before)
+    S = _streak_mask_directional(A, angle_deg=angle)
+
+    # anisotropic comet matte
+    M2D = make_comet_mask_anisotropic(
+        A, angle_deg=angle,
+        core_k=1.2, tail_boost=0.9,
+        exclude_streaks=S,
+        feather_long=max(70.0, 4.5*feather_px),   # long feather down the tail
+        feather_cross=float(feather_px)           # tight across the tail
+    )
+    M2D *= float(mix)
+
+    if ch == 3:
+        M = np.repeat(M2D[...,None], 3, axis=2)
+    else:
+        M = M2D
+
+    out = A * M + B * (1.0 - M)
+    return out.astype(np.float32, copy=False)
+
+
+
+time_key = _minmax_time_key 
+
+
+def _protect_core_mask(h: int, w: int, cx: float, cy: float, r: float, soft: float) -> np.ndarray:
+    """
+    Radial soft mask centered at (cx,cy): 1 near core (protected), 0 far.
+    r = hard radius, soft = feather (pixels).
+    Returns 2D float32 [0..1].
+    """
+    yy, xx = np.mgrid[0:h, 0:w].astype(np.float32)
+    d = np.hypot(xx - float(cx), yy - float(cy))
+    m = np.clip((r + soft - d) / max(1e-6, soft), 0.0, 1.0)
+    return m.astype(np.float32)
+
+def _starless_frame_for_comet(img: np.ndarray,
+                              tool: str,
+                              settings,
+                              *,
+                              is_linear: bool,
+                              core_mask: np.ndarray) -> np.ndarray:
+    """
+    Run selected remover on a single frame and protect the nucleus with core_mask (H,W).
+    Returns RGB float32 [0..1] starless, with nucleus restored from original.
+    """
+    # ensure RGB float32 [0..1]
+    if img.ndim == 2: src = np.stack([img]*3, axis=-1).astype(np.float32)
+    elif img.ndim == 3 and img.shape[2] == 1: src = np.repeat(img, 3, axis=2).astype(np.float32)
+    else: src = img.astype(np.float32, copy=False)
+
+    # run
+    if tool == "CosmicClarityDarkStar":
+        # DarkStar returns in the same domain we fed in.
+        base_for_mask = src
+        starless = darkstar_starless_from_array(src, settings)
+
+        # protect nucleus (blend original back where mask=1), in *current* domain
+        m = core_mask.astype(np.float32)
+        m3 = np.repeat(m[..., None], 3, axis=2)
+        protected = starless * (1.0 - m3) + base_for_mask * m3
+        return np.clip(protected, 0.0, 1.0)
+
+    else:
+        # StarNet path: do mask-blend inside the function (in its stretched domain)
+        protected, _ = starnet_starless_pair_from_array(
+            src, settings, is_linear=is_linear, core_mask=core_mask  # NOTE: keyword arg
+        )
+        return np.clip(protected, 0.0, 1.0)
+
+
+def _gamma_stretch(x: np.ndarray, gamma: float = 0.6,
+                   lo_pct: float = 1.0, hi_pct: float = 99.7) -> np.ndarray:
+    """
+    Percentile-clip → normalize to [0,1] → power-law gamma → back to float32.
+    gamma < 1 brightens midtones (good for faint coma).
+    """
+    x = np.asarray(x, dtype=np.float32)
+    lo = np.percentile(x, lo_pct)
+    hi = np.percentile(x, hi_pct)
+    if hi <= lo:
+        return x  # degenerate; skip
+    y = np.clip((x - lo) / (hi - lo), 0.0, 1.0)
+    y = np.power(y, gamma, dtype=np.float32)
+    return y
+
+def _refine_centroid(L: np.ndarray, px: float, py: float, r: int = 12) -> Tuple[float, float]:
+    """
+    Subpixel refinement around (px,py) using an intensity-weighted centroid
+    on a small ROI after subtracting a robust local background.
+    """
+    H, W = L.shape
+    x1 = max(0, int(round(px - r)));  x2 = min(W, int(round(px + r + 1)))
+    y1 = max(0, int(round(py - r)));  y2 = min(H, int(round(py + r + 1)))
+    roi = L[y1:y2, x1:x2].astype(np.float32, copy=False)
+    if roi.size < 16: 
+        return px, py
+
+    m = np.median(roi)
+    s = np.std(roi)
+    thr = m + 1.0 * s
+    w = roi - thr
+    w[w < 0] = 0.0  # keep only positive contrast (coma/core)
+    if not np.any(w):
+        return px, py
+
+    ys, xs = np.mgrid[y1:y2, x1:x2]
+    Wsum = float(w.sum())
+    cx = float((w * xs).sum() / Wsum)
+    cy = float((w * ys).sum() / Wsum)
+    return cx, cy
+
+
+
+# ---------------- Qt6-only centroid review dialog ----------------
+try:
+    # Prefer PyQt6
+    from PyQt6.QtCore import Qt, QPointF, QEvent
+    from PyQt6.QtGui import QImage, QPixmap, QPainter, QPen, QCursor
+    from PyQt6.QtWidgets import (
+        QDialog, QListWidget, QListWidgetItem, QLabel, QPushButton, QHBoxLayout,
+        QVBoxLayout, QSlider, QWidget, QSpinBox, QCheckBox, QGraphicsView,
+        QGraphicsScene, QGraphicsPixmapItem, QGraphicsEllipseItem
+    )
+    _QT_BINDING = "PyQt6"
+except Exception:
+    # Fallback to PySide6 (still Qt6)
+    from PySide6.QtCore import Qt, QPointF, QEvent
+    from PySide6.QtGui import QImage, QPixmap, QPainter, QPen
+    from PySide6.QtWidgets import (
+        QDialog, QListWidget, QListWidgetItem, QLabel, QPushButton, QHBoxLayout,
+        QVBoxLayout, QSlider, QWidget, QSpinBox, QCheckBox, QGraphicsView,
+        QGraphicsScene, QGraphicsPixmapItem, QGraphicsEllipseItem
+    )
+    _QT_BINDING = "PySide6"
+
+CursorShape = Qt.CursorShape 
+
+class CometCentroidPreview(QDialog):
+    """
+    Qt6 dialog to review/adjust comet centroids for a list of frames.
+    Returns { path: (x, y) } via get_seeds() after accept().
+    """
+    def __init__(self, file_list, initial_xy=None, parent=None):
+        super().__init__(parent)
+        self.setWindowTitle("Comet: Review & Adjust Centroids")
+        self.files = list(file_list)
+        self.xy = dict(initial_xy or {})
+        self.gamma = 0.6
+        self.blur  = 3.5
+        self.dot_r = 12
+        self.zoom  = 1.0
+
+        # --- left: list ---
+        self.listw = QListWidget()
+        for p in self.files:
+            it = QListWidgetItem(os.path.basename(p))
+            it.setToolTip(p)
+            self.listw.addItem(it)
+        self.listw.currentRowChanged.connect(self._on_select)
+
+        # --- center: graphics view ---
+        self.scene = QGraphicsScene(self)
+        self.view  = QGraphicsView(self.scene)
+        self.view.setRenderHints(
+            self.view.renderHints()
+            | QPainter.RenderHint.Antialiasing
+            | QPainter.RenderHint.SmoothPixmapTransform
+        )
+        self.view.setDragMode(QGraphicsView.DragMode.NoDrag)
+        self.view.setCursor(QCursor(CursorShape.ArrowCursor))
+        self.view.viewport().setCursor(QCursor(CursorShape.ArrowCursor))
+        self.pix_item = QGraphicsPixmapItem()
+        self.scene.addItem(self.pix_item)
+        self.cross = QGraphicsEllipseItem(-self.dot_r, -self.dot_r, 2*self.dot_r, 2*self.dot_r)
+        pen = QPen(Qt.GlobalColor.green); pen.setWidthF(1.5)
+        self.cross.setPen(pen)
+        self.scene.addItem(self.cross)
+        self.view.viewport().installEventFilter(self)
+
+        # --- right: controls ---
+        self.s_gamma = QSlider(Qt.Orientation.Horizontal); self._prep_slider(self.s_gamma, 10, 200, int(self.gamma*100))
+        self.s_blur  = QSlider(Qt.Orientation.Horizontal); self._prep_slider(self.s_blur, 0, 80, int(self.blur*10))
+        self.s_zoom  = QSlider(Qt.Orientation.Horizontal); self._prep_slider(self.s_zoom, 10, 300, int(self.zoom*100))
+        self.s_gamma.valueChanged.connect(self._refresh_current)
+        self.s_blur.valueChanged.connect(self._refresh_current)
+        self.s_zoom.valueChanged.connect(self._apply_zoom)
+
+        self.n_prop = QSpinBox(); self.n_prop.setRange(1, 50); self.n_prop.setValue(3)
+        self.cb_show_gamma = QCheckBox("Show gamma preview"); self.cb_show_gamma.setChecked(True)
+
+        self.btn_auto   = QPushButton("Auto")
+        self.btn_prev   = QPushButton("⟲ Prev")
+        self.btn_next   = QPushButton("Next ⟳")
+        self.btn_copyf  = QPushButton("Propagate →")
+        self.btn_ok     = QPushButton("OK")
+        self.btn_cancel = QPushButton("Cancel")
+
+        self.btn_auto.clicked.connect(self._auto_pick)
+        self.btn_prev.clicked.connect(lambda: self._change_row(-1))
+        self.btn_next.clicked.connect(lambda: self._change_row(+1))
+        self.btn_copyf.clicked.connect(self._propagate_forward)
+        self.btn_ok.clicked.connect(self.accept)
+        self.btn_cancel.clicked.connect(self.reject)
+
+        ctrls = QVBoxLayout()
+        ctrls.addWidget(QLabel("Gamma")); ctrls.addWidget(self.s_gamma)
+        ctrls.addWidget(QLabel("Blur σ")); ctrls.addWidget(self.s_blur)
+        ctrls.addWidget(QLabel("Zoom")); ctrls.addWidget(self.s_zoom)
+        ctrls.addWidget(self.cb_show_gamma)
+        r1 = QHBoxLayout(); r1.addWidget(self.btn_auto); r1.addWidget(self.btn_copyf); r1.addWidget(self.n_prop); ctrls.addLayout(r1)
+        r2 = QHBoxLayout(); r2.addWidget(self.btn_prev); r2.addWidget(self.btn_next); ctrls.addLayout(r2)
+        ctrls.addStretch(1)
+        r3 = QHBoxLayout(); r3.addWidget(self.btn_ok); r3.addWidget(self.btn_cancel); ctrls.addLayout(r3)
+
+        main = QHBoxLayout(self)
+        main.addWidget(self.listw, 1)
+        main.addWidget(self.view, 4)
+        w = QWidget(); w.setLayout(ctrls)
+        main.addWidget(w, 2)
+
+        self.cb_show_gamma.toggled.connect(self._refresh_current)
+
+        if self.files:
+            self.listw.setCurrentRow(0)
+
+        if self.files and self.files[0] not in self.xy:
+            self._auto_pick(one_file=self.files[0], silent=True)
+            self._place_cross()
+
+        self.view.viewport().installEventFilter(self)     
+
+    def eventFilter(self, obj, ev):
+        if obj is self.view.viewport():
+            if ev.type() == QEvent.Type.CursorChange:
+                obj.setCursor(QCursor(CursorShape.ArrowCursor))
+                return True
+            if ev.type() == QEvent.Type.MouseButtonPress:
+                if ev.button() == Qt.MouseButton.LeftButton:
+                    pos = self.view.mapToScene(ev.position().toPoint())
+                    self._set_xy_current(pos.x(), pos.y())
+                    return True
+        return super().eventFilter(obj, ev)
+
+
+    # --- Qt6 helpers ---
+    def _prep_slider(self, s, lo, hi, val):
+        s.setRange(lo, hi); s.setValue(val); s.setSingleStep(1); s.setPageStep(5)
+
+    def eventFilter(self, obj, ev):
+        if obj is self.view.viewport() and ev.type() == QEvent.Type.MouseButtonPress:
+            if ev.button() == Qt.MouseButton.LeftButton:
+                pos = self.view.mapToScene(ev.position().toPoint())
+                self._set_xy_current(pos.x(), pos.y())
+                return True
+        return super().eventFilter(obj, ev)
+
+    def keyPressEvent(self, ev):
+        k = ev.key()
+        if k in (Qt.Key.Key_Left, Qt.Key.Key_Right, Qt.Key.Key_Up, Qt.Key.Key_Down):
+            dx = -0.5 if k == Qt.Key.Key_Left else (0.5 if k == Qt.Key.Key_Right else 0.0)
+            dy = -0.5 if k == Qt.Key.Key_Up   else (0.5 if k == Qt.Key.Key_Down  else 0.0)
+            f = self._cur_file()
+            if f in self.xy:
+                x,y = self.xy[f]; self.xy[f] = (x+dx, y+dy); self._place_cross()
+            ev.accept(); return
+        super().keyPressEvent(ev)
+
+    # --- logic ---
+    def _cur_file(self):
+        r = self.listw.currentRow()
+        return self.files[r] if 0 <= r < len(self.files) else None
+
+    def _change_row(self, delta):
+        r = self.listw.currentRow()
+        self.listw.setCurrentRow(max(0, min(len(self.files)-1, r+delta)))
+
+    def _apply_zoom(self):
+        self.zoom = max(0.1, self.s_zoom.value()/100.0)
+        self.view.resetTransform()
+        self.view.scale(self.zoom, self.zoom)
+
+    def _render_preview(self, img):
+        if self.cb_show_gamma.isChecked():
+            sigma = max(0.0, self.s_blur.value()/10.0)
+            g = max(0.1, self.s_gamma.value()/100.0)
+            L = _luma_gauss(img, sigma if sigma>0 else 0.0)
+            G = _gamma_stretch(L, gamma=g)
+            disp = cv2.normalize(G, None, 0, 255, cv2.NORM_MINMAX).astype(np.uint8)
+        else:
+            L = _to_luma(img)
+            disp = cv2.normalize(L, None, 0, 255, cv2.NORM_MINMAX).astype(np.uint8)
+        return disp
+
+    def _on_select(self, row):
+        fp = self._cur_file()
+        if not fp: return
+        img, _, _, _ = load_image(fp)
+        if img is None: return
+        disp = self._render_preview(img)
+        qimg = QImage(disp.data, disp.shape[1], disp.shape[0], disp.strides[0], QImage.Format.Format_Grayscale8)
+        self.pix_item.setPixmap(QPixmap.fromImage(qimg.copy()))
+        self.scene.setSceneRect(0, 0, disp.shape[1], disp.shape[0])
+        if fp not in self.xy:
+            self._auto_pick(one_file=fp, silent=True)
+        self._place_cross()
+        self._apply_zoom()
+
+    def _place_cross(self):
+        fp = self._cur_file()
+        if not fp or fp not in self.xy: return
+        x,y = self.xy[fp]
+        self.cross.setPos(QPointF(x, y))
+
+    def _set_xy_current(self, x, y):
+        fp = self._cur_file()
+        if not fp: return
+        self.xy[fp] = (float(x), float(y))
+        self._place_cross()
+
+    def _auto_pick(self, one_file=None, silent=False):
+        targets = [one_file] if one_file else [self._cur_file()]
+        hint = max(4.0, self.s_blur.value()/10.0)
+        sigmas = [0.6*hint, 0.9*hint, 1.3*hint, 1.8*hint, 2.4*hint]
+
+        for fp in targets:
+            if not fp: continue
+            img, _, _, _ = load_image(fp)
+            if img is None: continue
+            L = _to_luma(img).astype(np.float32)
+
+            # 1) try local search around existing xy (seed or previous)
+            cx0, cy0 = self.xy.get(fp, (None, None))
+            found = False
+            if cx0 is not None:
+                half = max(24, int(3*hint))
+                x1,y1,x2,y2 = _crop_bounds(cx0, cy0, half, L.shape[1], L.shape[0])
+                Ls = _star_suppress(L[y1:y2, x1:x2])
+                cx, cy, used = _log_big_blob(Ls, sigmas)
+                cx += x1; cy += y1
+                g = max(0.1, self.s_gamma.value()/100.0)
+                cx, cy = _refine_centroid(_gamma_stretch(L, g), float(cx), float(cy), r=max(10, int(used)))
+                self.xy[fp] = (float(cx), float(cy))
+                found = True
+
+            # 2) global fallback
+            if not found:
+                Ls = _star_suppress(L)
+                cx, cy, used = _log_big_blob(Ls, sigmas)
+                g = max(0.1, self.s_gamma.value()/100.0)
+                cx, cy = _refine_centroid(_gamma_stretch(L, g), float(cx), float(cy), r=max(10, int(used)))
+                self.xy[fp] = (float(cx), float(cy))
+
+        self._place_cross()
+        if not silent:
+            self._refresh_current()
+
+    def _propagate_forward(self):
+        n = int(self.n_prop.value())
+        r = self.listw.currentRow()
+        if r < 0: return
+        fp = self.files[r]
+        if fp not in self.xy: return
+        for k in range(1, n+1):
+            i = r + k
+            if i >= len(self.files): break
+            self.xy[self.files[i]] = self.xy[fp]
+        self._change_row(+1)
+
+    def get_seeds(self):
+        return dict(self.xy)
+
+    def _refresh_current(self):
+        """Re-render current frame with the latest gamma/blur and keep the cross in place."""
+        r = self.listw.currentRow()
+        if r < 0 or r >= len(self.files):
+            return
+        fp = self.files[r]
+        img, _, _, _ = load_image(fp)
+        if img is None:
+            return
+        disp = self._render_preview(img)
+        qimg = QImage(disp.data, disp.shape[1], disp.shape[0], disp.strides[0],
+                    QImage.Format.Format_Grayscale8)
+        self.pix_item.setPixmap(QPixmap.fromImage(qimg.copy()))
+        self.scene.setSceneRect(0, 0, disp.shape[1], disp.shape[0])
+        self._place_cross()   # keep marker where it was