Autocrop finds the subject in a background-removed image

[JamesmarkeyUK]

Autocrop on a cut-out could hand back exactly the crop that was already there, looking like it had failed to see the picture at all. Reproduced in a headless browser, matching the reported rectangle byte for byte: a 4480×2520 cut-out with a 2128×2128 crop at x=1064, y=0 came back unchanged.

Why

Two faults in the content scan, both triggered by background removal:

• Invisible alpha residue. Any pixel with alpha ≥ 16 counted as content. A matting model leaves a wash of near-zero alpha where it was unsure — invisible against the checkerboard, but enough to stretch the detected box to the edges of whatever was scanned. The box then filled its region, so 1:1 sized the square to the region and returned it. In the repro, alpha-20 haze turned the true subject box {1542, 681, 1786×1431} into {201, 60, 3899×2302}.
• A single speck pinned the box. Bounds came from a raw min/max over every marked pixel, so one mote of dust or one stray pixel in a corner blew out the crop — four isolated pixels widened a 1786px-wide subject to 3604px.

Scanning was also confined to the current crop, so a crop drawn through the subject could hide part of it from the search.

What changed

• Alpha decides a cut-out. Colour sampling is actively wrong on a transparent image: pixels under transparency read as black, so a dark subject gets called "background". The alpha channel is the answer already, with the bar at 64 — above the residue, below the subject's anti-aliased rim.
• Specks are grouped and ignored. Marked pixels are grouped into blobs (run-based connected components, ~65 ms on a 40 MP source), and a blob is dropped only when it is tiny on both axes and holds under 0.5% of the content found. A hairline rule or a wire has real extent, so it is never at risk; if everything looks like a speck, everything is kept.
• Background colour is the median of the border, not the average of four corners, so a subject touching an edge no longer poisons the sample.
• The scan is always full-frame and replaces whatever crop is set. computeContentBounds loses its region parameter along with the last caller that passed one.
• Autocrop never silently no-ops. The finished rectangle is compared with the crop you already have (or the whole image, when you have none); when they match, the crop is left alone and the panel says why instead of looking like a broken button.

Verified

Scenario	Before	After
Cut-out + stale 2128² crop, 1:1	unchanged 2128²	{1542, 504, 1786×1786} — whole subject
Same, Max	unchanged 2128²	{1542, 681, 1786×1431} — exact subject
Crop drawn through the subject	no-op	finds the full subject
Full-bleed photo, 1:1	centred square	centred square (unchanged)
Full-bleed photo, Max / Keep ratio	whole-image crop	"Nothing to trim"
Flat image	arbitrary centred square	crop untouched + note

Seven opaque-image regressions still pass — logo on white, photo on a black card, subject touching an edge, scan with a dust speck, flat colour, thin tall feature, black subject on transparency. npm run typecheck and npm run build are clean.

---

Generated by Claude Code