fix(ui): whitespace-only and block-boundary selections no longer throw or create unanchorable annotations

[backnotprop]

Problem

Closes #881.

A double-click at a block boundary in a markdown plan or annotate session — just past the end of a heading, paragraph, list item, fence or blockquote — leaves a non-collapsed selection whose whole string is the break between two blocks ("\n"). web-highlighter's own pointer-end listener acts on it, and which of two symptoms you get depends only on where the browser put the range's end:

1. A garbage annotation. The library paints an invisible highlight and the toolbar opens on an empty quote. Acting on it stores an annotation with originalText: "\n" — invisible, counted in the annotation total, exported to the agent, and unanchorable on the next load.
2. An uncaught TypeError. When the end boundary is an element with endOffset >= childNodes.length (what Chromium reports at the end of the last block), the library resolves it to childNodes[offset] — undefined — and Cannot read properties of undefined (reading 'parentNode') escapes its listener.

Reproduction (before)

Built from this branch's base (origin/main 9bea44e) and driven in headless Chromium against a real annotate session (plannotator annotate fixture.md, fixture with paragraph / list / heading / fence / blockquote / table adjacent). 44 real mouse gestures: double-click just past the end of each block's last text run (dx 1/4/10 px), double-click below each block's text baseline, double-click in each inter-block gap; plus a deterministic whitespace-selection-then-Delete flow. page.on('pageerror') captured, and the live selection recorded at mouseup on the capture phase.

	before	after
gestures	44	44
uncaught TypeErrors	3	0
blank-quote selections that painted a highlight + opened the toolbar	21	0
valid selections that still opened the toolbar	15	15
whitespace selection → Delete → GET /api/draft	stores {"type":"DELETION","originalText":"\n",…}	no toolbar, no annotation, draft untouched

The three throwers were dblclick-past-end-P-block-8-dx{1,4,10} (the last paragraph), stack:

TypeError: Cannot read properties of undefined (reading 'parentNode')
    at getOriginParent … at a.getDomMeta … at k.serialize
    at B.H._highlightFromHRange … at B.H._handleSelection

Evidence (screenshots, per-gesture JSON, drivers) in the scratchpad:
881/before-final/ (incl. persist-b-after.png: the panel showing Annotations 1 / a Deletion card with an empty quote, and dblclick-past-end-P-block-8-dx*.png), 881/after-final/, 881/repro6.mjs (the sweep), 881/repro3.mjs, 881/isolate-block5.mjs, 881/fixture.md.

Root cause

All in @plannotator/web-highlighter@0.8.1, reached from our own call sites:

• src/model/range/selection.ts getDomRange() refuses only a collapsed selection, so a whitespace-only one goes through.
• src/model/range/dom.ts formatDomNode() returns { $node: n.$node.childNodes[n.offset], offset: 0 } with no bounds check; offset === childNodes.length is legal in the DOM, so $node is undefined.
• getDomMeta → getOriginParent then reads $node.parentNode and throws, from inside the bubble-phase mouseup listener Highlighter.run() registers on $root — i.e. as an uncaught error, with no annotation created either way.
• Restore (applyAnnotationsInternal, packages/ui/hooks/useAnnotationHighlighter.ts) does try fromStore inside a try, but a blank originalText still "succeeds": the stored positions resolve, the content check compares whitespace against whitespace, and an invisible mark is painted on a spot the reviewer never chose — or the text fallback searches the DOM for "\n".

The fix belongs in this repo rather than the fork so it does not wait on a publish; the fork could additionally bounds-check formatDomNode and trim in getDomRange, and these guards would stay correct if it did.

Change

packages/ui/hooks/useAnnotationHighlighter.ts only (+109 lines, no behavior change for valid selections):

• Two predicates beside the existing compactText: isBlankQuote (whitespace-only, nbsp included) and isSerializableRange / boundaryResolves (a non-text boundary the library would resolve to undefined).
• handlePointerEndCapture: a blank or unserializable selection is collapsed and dropped. This runs on the capture phase of the same element, registered before highlighter.run(), so the library's own handler sees a collapsed selection and no-ops. It is deliberately placed ahead of the existing containment check: the library does not check containment either, and the last-block shape has its common ancestor in the scroll wrapper outside the article — bailing on containment first is what let those three throws through in my first pass.
• CREATE handler: a blank source.text (after the existing exclusion-aware quote repair, which can itself empty it) removes the painted highlight and returns — no toolbar, no composer, no quick-label, previous pending untouched. This covers every producer: the library's listener, the touch selectionchange bridge, vim and pinpoint ranges.
• createAnnotationFromSource: refuses a blank quote as a last line.
• Our own fromRange calls (highlightRange, the touch bridge) are wrapped in try/catch with a once-per-load console.warn; state is left untouched.
• Restore: a stored row with a blank originalText is skipped before attempted — neither painted nor reported, the same treatment a diagramAnchor row gets, so a legacy draft carrying one does not raise an "Unanchored" chip and a toast on every load.

Valid selections keep their exact anchoring and quote (pinned below); share/draft round-trips are untouched.

Tests

New DOM-gated file packages/ui/hooks/useAnnotationHighlighter.blankSelection.test.tsx (6 tests), added to the DOM_TESTS step in .github/workflows/test.yml (scripts/dom-test-allowlist.test.ts passes). It drives the real library over a real mouseup in happy-dom, where both shapes reproduce, and captures uncaught errors via window.addEventListener('error'):

• whitespace-only boundary selection → no annotation, no toolbar/composer/picker, no <mark> (before: two empty marks + composer);
• the same in redline mode → no annotation (before: one stored originalText: "\n");
• endOffset === childNodes.length → no uncaught error (before: the parentNode TypeError);
• a selection ending in an empty element outside the container (the real last-block shape) → no uncaught error;
• regression pin: a normal whole-paragraph selection still yields the identical annotation — originalText, blockId, offsets, and web-highlighter's startMeta/endMeta — and paints the same mark;
• restore: a stored originalText: "\n" row is skipped (not attempted, not unanchored, not painted) while a valid neighbour restores normally.

With the source change reverted, 5 of the 6 fail; the one that passes is the regression pin.

Counts on this branch:

• bun test packages/ui packages/editor → 1261 pass, 905 skip, 0 fail (2166 tests / 225 files)
• DOM_TESTS=1 bun test packages/ui packages/editor → 2154 pass, 1 skip, 4 fail — the 4 are the known pre-existing DocBadges open-in placement failures, unchanged from main
• typecheck → clean for core / shared / ai / server / ui / guides-show / strict-consumer / pi-extension. (packages/guide-viewer reports TS2688: Cannot find type definition file for 'bun' in my sandbox both with and without this change — an artifact of invoking tsc directly there, not from this PR.)

After

Rebuilt (bun run --cwd apps/review build && bun run build:hook) and re-ran the identical 44-gesture sweep against a fresh annotate server: 0 uncaught errors, 0 blank-quote highlights, the same 15 valid toolbars, and the whitespace-selection-then-Delete flow leaves the panel at "No annotations yet" with an untouched draft (881/after-final/persist-b-after.png vs 881/before-final/persist-b-after.png).

One gesture in the after run still shows a toolbar with a blank selection string: verified in isolation (881/isolate-block5.mjs) to be the unrelated code-block click feature — it creates a codeblock-… annotation quoting the fence's real source, and painted no selection highlight.