fix(#745): performance.spec.ts asserted nothing — T167 never paginated

[TortoiseWolfe]

performance.spec.ts had ten tests, nine of which asserted only that the thread and input
were visible, then waited a few animation frames and closed the page. T167: Pagination loads next 50 messages never paginated. Two tests claimed 60fps while doing no scrolling and taking
no timings. They passed whether the features worked, were broken, or were deleted.

Rewritten to six tests that each name an observable consequence. They immediately found
three defects, which is why this PR sat red — and all three are now resolved.

What the new tests found

1. The jump button was decided against an empty thread — fixed in #754/#755, already on
main.

2. The virtualized jump asked for a scroll TanStack cannot perform — also #754/#755.

3. The jump resolves to a coordinate while the page is still loading. Reaching the top is
what triggers loading older messages, so "scroll to top, then press jump" is the ordinary
case. The click resolved to scrollTop = scrollHeight at that instant — a number already
stale by the time the scroll finished — and nothing re-aimed, because prepending older
messages does not change the newest message id, so the auto-scroll effect never fires. Landed
1934px short in 2 of 3 full-spec runs.

Fixed by making the jump an intent that outlives the trip: while it is pending the
ResizeObserver re-aims at the bottom as content grows, and the reader taking over — wheel,
touch, keydown, scrollbar mousedown — cancels it. Deliberately not the scroll event, which
the jump's own animation emits and which would cancel the intent the moment it started acting
on it.

That observer also watched firstElementChild, which during pagination is the loader, not
the message list — so it would have missed exactly the growth this depends on. It now observes
every child and re-subscribes when content changes.

One correction

I reported loading older history does not teleport you to the top as a product defect. It
was my test measuring the gap. Traced in the browser:

  0ms  top=0     h=5278  kids=2
 94ms  top=2722  h=8216  kids=8
155ms  top=2962  h=8600  kids=14   ← stable for the next 1.4s

Restoration works. The test polled scrollHeight > before.scrollHeight where before was
5216 while the settled height was 5278, so 62px of layout drift satisfied the poll before any
older message arrived, and the one-shot read that followed caught scrollTop still at 0. It now
polls the value it claims, and separately asserts the thread really grew — so it cannot be
satisfied by a stray scroll with no pagination behind it.

Verification

Measured where it actually fails, which is the full spec — running the test alone passes
regardless, which is how the first mutation attempt fooled me:

	before	after
firefox, full spec	2 of 3 failing	6/6 clean
chromium, full spec	—	7/7 clean
webkit	cannot launch locally (missing libxslt, libx264) — CI covers it

Reverting the re-aim reproduces the failure. Two deterministic unit tests were added because
the E2E is timing-dependent by nature: one grows the content and drives the observer by hand,
asserting the jump re-aims at the new height; one asserts a wheel event stops it so it cannot
fight the reader. Each fails under its own mutation, and the first fails first if the click
never reached the container.

MessageThread units: 35/35.

Also filed

#757 — /game/cod-skeleton horizontal overflow failed once in CI and is 5/5 clean
locally. Recorded as unreproduced rather than dismissed as a flake, with the dev-server vs
production-build caveat stated. Not blocking.

Closes #745
Closes #756