fix(clone): drop <picture> <source> elements so the inlined <img> src is not overridden (#462)

[dylan1951]

Closes #462.

Problem

A <picture>'s <source> elements out-rank the <img>'s own src. They were cloned verbatim, so inside the exported SVG the picture selection algorithm re-selected an external URL — and SVG-as-image cannot load external resources. The photo never painted, however correctly the <img> had been inlined.

The signature is a large SVG that rasterizes to a tiny PNG: everything embedded, nothing drawn.

Why the existing machinery doesn't cover it

• freezeImgSrcset (src/utils/clone.helpers.js) already sets the cloned <img>'s src = original.currentSrc and strips srcset/sizes — but only on the <img>. The sibling <source> elements go through the generic child-clone path with their external srcsets intact.
• pictureResolver (src/modules/pictureResolver.js) only acts on a <picture> whose <img src> is a lazy-load placeholder (isPlaceholderSrc); it continues for a normally-loaded picture, and runs against the live DOM before cloning.

So a fully-loaded <picture> reaches the SVG with its <source> list intact.

Fix

Skip <source> children of a <picture> in deepClone:

if (tag === 'source' && node.parentElement?.localName === 'picture') {
  return null
}

Because freezeImgSrcset has already frozen the clone to the variant the live page chose, the sources carry nothing still needed. Art direction is preserved — the chosen crop comes from currentSrc, not from re-evaluating the sources. Capturing the same carousel at a 1600px viewport embeds the 2200×1200 crops; at 500px it embeds the 600×1000 crops.

The check is deliberately narrow: <source> inside <audio>/<video> is untouched, and there's a test covering that.

Results

Real WordPress hero carousel, real assets, identical 723 KB SVG both runs:

	<source> in SVG	embedded images	PNG	photo painted
before	4	229 KB + 280 KB	48 KB	0 %
after	0	229 KB + 280 KB	2046 KB	100 %

Tests

New deepClone — <picture> sources block in __tests__/core.clone.test.js:

• <source> children of a <picture> are dropped and the inlined <img> src survives — fails on main, passes with the fix
• a <source> outside a <picture> is kept — guards against over-reach; passes either way by design

Full suite: 685 passed / 1 skipped, against 683 / 1 on main — no regressions. npm run lint and npm run test:types clean. Verified on Chromium; I don't have the Firefox/WebKit Playwright binaries locally.

[tinchox5]

Merged, thanks. Good instinct scoping the fix to <picture>'s <source> specifically instead of touching <source> generally — the added test guarding <audio>/<video> sources is exactly the kind of thing that keeps a fix from becoming an over-reach.