[latekvo]

Closes #807

The 1% aspect tolerance let full-res captures from two different iPhones (16 vs 16 Pro / Pro Max, 0.29% / 0.22% apart) be resampled and diffed as one screen. The tolerance is now capped at 2 / shortest side, the most that integer rounding of rescales can move the aspect, so pairs further apart than that become a dimension mismatch while real 1.0x vs 0.5/0.3/0.25x pairs still normalize.

Pairs inside the rounding bound still normalize: they cannot be told apart from a rescale by dimensions alone (i16 Pro vs Pro Max, i16 vs 16 Plus, and more once one side is a scaled capture, since the bound grows as the image shrinks).

On a foldable, a foreign baseline near a panel aspect now hard-fails; the posture hint is dropped when both sizes map to one posture, instead of naming the cover panel twice and saying to fold.

No docs update needed.

Verification

Probe: diffPngFiles on real PNGs, with same-device downscales made by real sips -Z (the path downscalePngInPlace uses) and full-res cross-device pairs that sit within 1% aspect.

pair	main	this PR
21 same-device 1.0 vs sips 0.25/0.3/0.5 (7 devices)	NORMALIZED	NORMALIZED
i16 1179x2556 vs i16 Pro 1206x2622 (0.285%)	NORMALIZED	MISMATCH
i16 vs i16 Pro Max 1320x2868 (0.220%)	NORMALIZED	MISMATCH
i16 vs i16e 1170x2532 (0.177%)	NORMALIZED	MISMATCH
i16 Pro vs i16e / Pro Max vs i16e	NORMALIZED	MISMATCH
Pixel 9 1080x2424 vs Pixel 7 1080x2400 (0.990%)	NORMALIZED	MISMATCH
i16 Pro vs i16 Pro Max (0.064%)	NORMALIZED	NORMALIZED

The last row cannot be decided from dimensions: 1206x2622 is exactly a floor-rounded 0.914x rescale of 1320x2868. Images with a shortest side under 200 px keep the 1% bound, as before (100x200 vs 101x200 normalizes, vs 102x200 mismatches).

A sweep over 29,796 rescale pairs (13 device sizes, scales 0.05-1.0, round/floor/ceil) found 0 pairs that normalize on main and are rejected here. Reverting only the return line brings back both failures in the unit-level repro.

A second sweep modelled each producer's rounding: sim-server f32 round (iOS sim and Android), Math.round (Chromium, Vega, Harmony) and measured sips -Z (physical iOS, tvOS). It covered 80 device sizes, scaled-vs-scaled pairs (such as the 0.5 baseline vs 0.25 live-capture fallback) and double rescales, and found 0 same-screen pairs rejected here. iPhone 16 Pro Max simulator captures at 7 scales came out at exactly round(W*s). Mutants 2 -> 1 (now killed by the two-rescale test) and 2 -> 3/4, min -> max are all caught.

Checks: npx tsc --build, the test typecheck, vitest on the screenshot-diff, foldable and flow snapshot/visual test files, and prettier + eslint on the changed files all pass.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

_{Comment @coderabbitai help to get the list of available commands.}

[latekvo]

Superseded by #1403, folded into #1753: it removes the aspect-ratio tolerance this PR capped, so #807's pair (1179x2556 vs 1206x2622) is now a dimension_mismatch.