test: cover what core.ts actually produces

[l2ysho]

Third mutant-killing PR from the report at l2ysho.github.io/afpp.

The gap

Every existing core.ts test asserted the page count and stopped:

const data = await pdf2image(input);
assert.equal(data.length, 9);

Nine items came back, so the test passes — whatever is in them. 64 of the 193 mutants survived.

The tests

Text is joined with single spaces. Pinning the start of page one kills the mutants that empty the separator (parts.join(' ') → parts.join('')) or seed the parts array with junk.

Rendered pages honour the scale. Page one is 595x841 at scale 1 and 1190x1683 at scale 2. Both numbers sit in the png IHDR header, so no image library is needed. This kills getViewport({ scale }) → getViewport({}), which silently falls back to scale 1.

A pooled canvas is resized before reuse. test/example-page-sizes.pdf is a hand-written two-page PDF whose pages are 200x100 and 100x300. The first page's canvas goes back to the pool and is handed to the second, so dropping the reset call produces a page rendered at the wrong size. 682 bytes of plain ASCII.

Pages without text are rendered, not emptied. In mixed mode a page with no text items is rasterised and handed to the callback as a Buffer; a text page arrives as a string. example-img.pdf gives nine buffers, example.pdf nine strings. Nothing covered that split.

The callback guard works. A missing callback and a non-function callback both reject.

The scale bounds are accepted, not just their outside rejected. The suite rejected 0, -1, 0.01, 100 and NaN, but never checked that 0.1 and 10 are allowed — so scale < 0.1 could become scale <= 0.1 unnoticed. Folded into the existing scale validation block.

Result

	Before	After
src/core.ts	64.55 %	74.60 %
All files	67.92 %	70.94 %
Survived in core.ts	64	47

Why I stopped at 47

The rest is not worth a test. Broken down:

• 13 mutants pick between the pooled canvas factory and pdf.js's own. Both render identical output.
• 10 mutants are page.cleanup(), canvasFactory.destroy() and finally blocks. Resource handling with no observable result.
• 10 mutants are canvas pool internals — pool size caps and reuse checks. Performance behaviour, same output either way.
• 3 mutants replace Array.from({ length: numPages }) with Array.from({}). The array is filled by index afterwards, so it is equivalent.
• The remainder are equivalent or need a machine with more than 8 cores to observe.

Killing these means exporting internals or asserting on things no caller can see. That trades real test quality for a number.