fix(desktop): anchor scrollback prepends in content space

[tlongwell-block]

Summary

• Store message scroll anchors in content space (row top relative to the scrolled content element, measured rect-to-rect) instead of viewport space.
• Keep a viewport-space reading only as a native-scroll-anchoring safety check so the JS re-pin cannot double-apply a prepend the browser already absorbed.
• Update the smooth-scroll anchor path to preserve the content-space anchor instead of a stale viewport target.

Single file: desktop/src/features/messages/ui/useAnchoredScroll.ts (+75/−16). No flag — this is a straight bug fix to the existing writer, not a behavior experiment.

Why

During scrollback pagination, the older-history page commits on the deferred snapshot a few frames after the last scroll event. The anchor saved rect.top − containerTop in viewport space at that last scroll event; momentum kept scrolling during the gap; the prepend re-pin then scrollBy()-ed back to the stale offset — snapping the reading row down by exactly the distance scrolled in between, once per ~50-message page boundary. That is the reported "intermittent stutter down when scrolling up."

Content space is scroll-invariant by construction: wheel motion translates content and row together, so measured drift is only layout growth above the anchor — the compensation can no longer rewind a live fling. Rect-to-rect against the content element (not viewport + scrollTop) deliberately reads a single layout clock; WebKit's committed scrollTop desyncs from getBoundingClientRect mid-momentum (the Lane C frozen-offset family), and mixing the two clocks re-introduced ~17px of false drift.

Evidence (mid-fling-prepend probe)

New instrument: fling through real windowed pagination (600-row history served through getChannelWindowEvents, so fetchOlder genuinely fires mid-momentum), per-frame anchor sampling. Every prior instrument — including the portable classifier — used fully pre-loaded corpora where fetchOlder never fires, which is why they scored clean on builds that still stuttered.

build	chromium	webkit
main (= #1678 flag-off)	jerk 17–20px at every prepend	jerk 17–20px at every prepend
this fix	12 prepends, 1 jerk (18px)	12 prepends, 3 jerks (max 20px)
this fix + #1678 eager layout	12 prepends, 1 jerk	11 prepends, 0 jerks

The residual jerks are a different write class: content-visibility re-realization above the anchor at commit time — exactly what #1678 (eagerTimelineLayout) removes. The two changes compose: this fix kills the stale-offset rewind (the per-page-boundary stutter), #1678 kills realization drift. All numbers repeatable across runs; probe run against this exact tree (f46ebe2c).

Validation

• pnpm -C desktop test — 2266 passed
• pnpm -C desktop typecheck / lint — clean (one pre-existing useOptionalChain warning in useAutoRestartPolicy.ts, untouched)
• pnpm -C desktop build — passed
• Playwright smoke on this tree: timeline-no-shift + scroll-history + overscroll-boundary + channel-window-mock-paging — 24/24
• messaging.spec.ts — 29/31; the 2 fails (:164 link previews, :533 avatar attribute) reproduce identically on the base commit 46613dcc with this change absent — pre-existing, not introduced here
• Mid-fling probe both engines, fix-only and composed with feat(desktop): eagerTimelineLayout preview — lay out fetched rows immediately (no content-visibility skipping) #1678 (table above)

Notes

• The probe fixture (upscroll-midfling-prepend.perf.ts) is deliberately not committed here to keep the fix standalone; it lives in the perf-lane worktrees and can land with the instrument set if wanted.
• Complementary to draft feat(desktop): eagerTimelineLayout preview — lay out fetched rows immediately (no content-visibility skipping) #1678 (eager timeline layout) and Polled settle gate for the mid-history prepend re-anchor (default-off) #1676 (settle gate); this one needs no flag and addresses the write class those two cannot: compensation aimed at a stale offset.