fix: render real-size partial images at the document-to-pixel scale

[palmoni5]

Problem

A custom layoutPages that does not lay pages out at their natural point size gets a permanently blurry page when the user zooms in — the high-resolution partial render never reaches the resolution the viewer is actually displaying, and zooming further does not help.

Cause

In _paintPagesCustom:

final pageScale = scale * max(rect.width / page.width, rect.height / page.height);
if (!enableLowResolutionPagePreview || pageScale > previewScaleLimit) {
  _requestRealSizePartialImage(cache, page, pageScale, targetRect);
}

scale is _currentZoom * devicePixelRatio — the document-units-to-physical-pixels factor. pageScale multiplies it by the layout ratio r = pageRect.size / page.size, which is the right quantity for the previewScaleLimit comparison, because previewScaleLimit is a page-points-to-pixels density (the preview is rendered as page.width * previewScaleLimit).

But _createRealSizePartialImage then uses the value it receives as a document-to-pixel factor again:

final width = (inPageRect.width * scale).toInt();   // inPageRect is in document units
fullWidth: pageRect.width * scale                   // pageRect is in document units

So r is counted twice:

needed:   pageRect.width * zoom * dpr
rendered: pageRect.width * (zoom * dpr * r)

With the built-in layouts r == 1, so the two are equal and nothing is wrong — which is why this only shows up with a custom layoutPages. With r = 0.5 (a two-page book-style spread laid out at half size, which is how I hit this) the partial image is rendered at half the required resolution and is upscaled 2x on screen. The error is proportional, so it never improves with more zoom.

Fix

Pass scale rather than pageScale to _requestRealSizePartialImage. The threshold comparison keeps using pageScale, which is correct there.

This is a no-op for every layout with r == 1, i.e. all built-in layouts.

Testing

To be explicit about what I have and have not checked: the analysis above is derived from reading the code, and I have not yet run a visual before/after in a built app. What I can state is that the change is provably a no-op wherever r == 1 (pageScale and scale are equal there by definition), which covers every built-in layout, so the risk to existing users is nil.

Happy to add a regression test if you can point me at where you would want it — the value is computed inside a private paint method, so I did not see an obvious seam in packages/pdfrx/test.