fix(desktop): keep history prepend anchored on the visible row

[tlongwell-block]

The bug

Scrolling up in a channel to load older history yanks the viewport to the
oldest just-loaded message instead of keeping the row the user was reading
pinned to the same screen position. Recurring regression; previous "fixes"
relied on math against container.scrollTop, which is stale on macOS
WKWebView mid-wheel (root cause documented in
RESEARCH/BUZZ_SCROLL_PREPEND_2026-06-14.md and Element/Matrix
docs/scrolling.md).

The fix (two parts, in useLoadOlderOnScroll.ts + MessageTimeline.tsx)

1. Anchor on a real visible DOM node, not on a scrollTop value.
   When the top sentinel intersects, findVisibleMessageAnchor walks
   [data-message-id] rows and picks the one whose top-offset is closest
   to min(12px, containerHeight / 2) — the row the user is most likely
   reading. We snapshot { id, topOffset, fallbackScrollHeight } in a
   ref while we wait for the prepended rows to commit.

2. Restore via relative scrollBy(delta), no scrollTop read.
   In a layout-effect keyed on the actual rendered first-id + count
   changing (i.e., the prepend has committed), we read the anchor row's
   new top-offset and call restoreScrollBy(nextTopOffset - savedTopOffset).
   This is the difference between "passes headless Chromium" and
   "actually fixed on macOS" — no stale read, no read-then-write.

MessageTimeline.tsx switches the consumer from restoreScrollPosition
(absolute write) to restoreScrollBy (relative), and changes the
fetch-older spinner from an in-flow block to an absolute overlay so the
spinner's own mount doesn't shove the topmost message down and confuse
the anchor math.

useTimelineScrollManager.ts adds the restoreScrollBy primitive
(container.scrollBy(0, delta)) alongside the existing scroll API.

The regression gate

Mock-bridge invariant test in messaging.spec.ts ("older-history
prepend keeps the visible row pinned"):

• Seeds a history-jump mock channel with 240 messages — enough to
  force at least one older-fetch round-trip past the 200-msg initial
  load.
• New historyDelayMs config knob delays only the older-history REQ
  with until (not initial loads), giving a deterministic snapshot
  window after the sentinel fires but before the prepend commits.
• Uses real mouse.wheel events, snapshots the anchor row by
  [data-message-id] rect against the timeline container, awaits
  prepend commit via first-id-changed + count-grew, then asserts
  ≤ 4px drift after layout-effect restore.
• Accounts for scrollTop drift between snapshots so it doesn't
  false-pass when scrollBy is wired correctly but the user kept
  wheeling.

A nightly-gated scroll-prepend-stability.spec.ts is included for
manual real-relay smoke (excluded from both smoke and integration
testMatch in playwright.config.ts). Run manually with
pnpm -C desktop exec playwright test scroll-prepend-stability.spec.ts
against a relay started with BUZZ_RECONCILE_CHANNELS=1.

Verified

• Positive test green against the fixed build (~3s, single worker).
• Negative test (restore neutered to void delta) fails as expected —
  confirms the gate catches the regression.
• pnpm typecheck + pnpm check (biome) clean.
• Build clean.

What's NOT done in this PR

• Manual macOS wheel-scroll repro. Headless Chromium ≠ WKWebView.
  A green CI run here is necessary, not sufficient. @tlongwell-block —
  pulling this branch on macOS and confirming the jump is gone is the
  load-bearing pass before this is "shipped."
• Empty-older-page guard. If fetchOlder resolves with zero new rows,
  hasCommittedPrepend never goes true, the layout effect early-returns
  leaving pendingRestoreRef set, and the observer never re-arms — older
  loading silently wedges. Happy path is covered; this is a follow-up
  commit on the same PR (resume observing if the fetch resolves with no
  commit). Landed as da1606c2 — converts the fetch handler to
  .then(onFulfilled, onRejected) and uses a single-rAF settle fence with
  an identity check on the captured pending to re-arm only on a true
  no-commit, never racing a real commit's restore.