fix(viewer): keep deep-linked inline comment thread aligned to its highlight

[yumike]

Problem

Opening an inline-comment deep link (#comment-<id>) could leave the comment thread in the right-margin column pinned hundreds of pixels above its highlighted passage. The thread is positioned by an article-relative offset (activeTop) that only recomputed on the article's ResizeObserver — which reports the article's box size, not the highlight's position within it. So when content above the highlight reflowed (a web-font swap on cold load, or a late image/diagram), the highlight slid down without the article resizing, the offset went stale, and the thread floated away from its passage.

A second symptom on hard refresh: the thread appeared instantly but mispositioned, then snapped into place ~1s later.

Fix

Trigger the re-measure at the right altitude, with two complementary mechanisms:

• observeMove — a small IntersectionObserver primitive (dynamic rootMargin, re-arming on each move; a dependency-free MIT port of Floating UI's autoUpdate({layoutShift})) that fires when an anchored element moves even though its own box is unchanged. Exposed as useElementMove, wired into the margin pin (active highlight) and useAnchorOffset (so the narrow-screen comment popover re-aligns too). Catches any reflow after the thread exists — live-reload re-wraps, late images/diagrams. The move getter reads comments.items so it re-subscribes to the fresh wrapper after a reconcile rebuilds highlights.
• A one-shot re-measure right after a deep-link reveal. On a cold load the stranding reflow happens during the load burst — before observeMove is watching, and while the highlight is still off-screen (so observeMove would only correct it via its ~1s poll). By reveal time the reflow has settled, so a single re-measure there aligns it immediately instead of after a visible ~1s delay.

Verification

• Reproduced on a real long page: deep link landed −683px off on the old build → −39px aligned on the fixed build; the hard-refresh settle dropped from ~1000ms → ~3ms (1 frame), consistently across reloads.
• Generalization confirmed: a simulated late image/diagram above the highlight re-aligns within ~2 frames (the case the article ResizeObserver structurally can't see).
• 591 viewer unit tests (new observeMove/useElementMove/useAnchorOffset-move coverage), 45 comment e2e, typecheck + lint all green; no regression in existing deep-link/alignment tests.

Follow-up (not in this PR)

The upstream trigger is unmanaged web-font FOUT (fonts.css ships font-display: swap with no metric-matched fallback) plus dimensionless <img>. Adding size-adjust/ascent-override fallback faces is an independent CLS win that would shrink the deep-link settle to a rare belt-and-suspenders path — tracked separately.

🤖 Generated with Claude Code