Windowed mounting by default: memory tracks the visible window, not assignment length

[dqnykamp]

Fixes #35, fixes #36, and fixes #37. Requires @doenet/doenetml-iframe ≥ 0.7.20-dev.332 (windowed mounting with keepLive and report-callback host support, Doenet/DoenetML#1479).

Every item's viewer component stays mounted, but which items are booted is now governed by the iframe wrapper's windowed mounting policy:

• An item only creates its iframe (a multi-MB standalone-bundle parse plus a core worker) when it comes near the viewport or into the pagination window — an item the student never reaches never boots at all (Mount DoenetViewer iframes only when rendered — every document currently creates its iframe (and 29 MB bundle load) immediately #35), and simultaneous boots are capped page-wide.
• At most maxLiveViewers (default 3) stay live; the rest are parked losslessly — their state is flushed before the iframe is detached, and they are restored (typed work intact, no interaction) when the student returns. A 20-question assignment paged all the way through holds ~3 live viewers instead of 20 (Paginated mode: bound mounted viewers to a window (current ± 1) and unmount items outside it #36); scroll-mode memory tracks the viewport (Scroll mode: shrink the 1000px IntersectionObserver rootMargin and unmount items that leave the viewport #37).
• In paginated mode the current page and its neighbors are marked keepLive, so page flips within the window stay instant while hidden pages beyond it never boot.

The viewer's own render-scheduling machinery (itemsToRender/itemsVisible/checkRender/per-item IntersectionObserver) is deleted in favor of the wrapper's policy. Parking requires flags.allowSaveState or allowLocalState; without a persistence path viewers still mount lazily but stay live once booted, so no work can be lost.

New ActivityViewer props:

• mountPolicy — overrides for maxLiveViewers / visibleMargin / parkDelayMs / flushTimeoutMs / maxConcurrentBoots;
• useSharedCoreWorker — serve all documents' cores from a shared worker pool instead of one dedicated ~100 MB worker per document (default off);
• working standaloneUrl / cssUrl / doenetmlVersion passthroughs — the standalone-bundle URL, its CSS, and a version override for every document's viewer;
• doenetViewerUrl and doenetMediaUrl are threaded through Viewer → Activity → SingleDocActivity to each embedded DoenetViewer. Both are inner-viewer props (forwarded across the iframe boundary), not bundle URLs: doenetViewerUrl is the <ref> renderer's activity-link base (it was not being forwarded correctly), and doenetMediaUrl is the <image source="doenet:…"> media base ([Feature]: Expose doenetMediaUrl in assignment-viewer package DoenetML#1457), newly exposed so doenet: images resolve.

Tests

New cypress spec (against the real published dev-channel bundle) covering both modes end to end:

• Paginated (5 docs): current+neighbor boot, pages beyond the window never do, iframe count stays ≤ 3; typing on page 1, paging to 4 (page 1 parks — iframe detached), and returning restores the typed work with no interaction.
• Scroll (5 docs, budget 2): off-screen items never boot; scrolling boots the far items and parks the ones left behind (iframe count ≤ 2).

A separate spec asserts that doenetViewerUrl and doenetMediaUrl reach the embedded viewer — both are baked into the booted DoenetViewer's iframe srcdoc, confirming they thread all the way through.

The park/restore round trip also exercises DoenetML#1479's report-callback capture path in a real host. Full suite green (47 vitest + 14 cypress), lint/build clean.

🤖 Generated with Claude Code