fix(conversation): make the top scroll FAB an overlay instead of a blocker

[RonenMars]

Problem

The ↑ Top FAB in the conversation detail view appears once you scroll past ~200px, horizontally centered at the top of the list, and sits on top of message content.
It was fully opaque, so text behind it was unreadable, and it intercepted touches in its rect.
Message bubbles span nearly the full width and the pill is centered, so it lands squarely on the text column rather than in a margin.

Occlusions captured during the #470 investigation: …write the spec doc for th[↑ Top]n plan. and | Cypress | Testing | s[↑ Top]ntend | strong |.

Root cause

ConversationHistoryList was the odd one out among three sibling implementations of the same affordance.

	background	show/hide	wrapper
ConversationHistoryList (before)	theme.text.accent, opaque	conditional mount/unmount, no fade	none
ConversationList	rgba(31, 111, 235, 0.14) + rgba(88, 166, 255, 0.2) border	Animated.View + useAnimatedStyle opacity, 220ms withTiming	pointerEvents="box-none"
TerminalOutput	rgba(31, 111, 235, 0.18)	same animated pattern	pointerEvents="box-none"

The two working versions use a translucent chip with muted text that reads as an overlay you can see through, and they fade rather than pop.
ConversationHistoryList used a solid accent-blue pill with white bold text, which reads as a modal blocker.

Fix

ConversationHistoryList now matches its two siblings rather than introducing a third visual language: translucent background, muted 12px text, Animated.View + pointerEvents="box-none" wrapper, 220ms fade in/out.
The show/hide condition moves from a useState toggle to a shared value, so the pill is always mounted and fades instead of popping.
pointerEvents tracks opacity inside the same animated style, so a faded-out pill cannot intercept list touches.

This is fixed purely as an overlay.
There are no contentContainerStyle padding, drawDistance, maintainVisibleContentPosition, or getItemType changes — those lines are load-bearing for the FlashList v2 anchoring stabilized in #470, which measured content-height swings up to ±19,898pt causing viewport teleports.
Padding that toggles with FAB visibility would mutate content height mid-scroll and reintroduce that class of bug.

The bottom CaretDown FAB is untouched; it sits bottom-right in the gutter and was not reported.

Before / after

Same fixture conversation, same scroll offset, driven by an identical scripted gesture sequence on a booted simulator against two Release builds.
In the before frame the table row reads | Express | Language | s[↑ Top] backend |.
In the after frame the same row reads | Express | Language | senior | backend | straight through the pill.

Before	After

Accessibility labels

The top FAB's accessibilityLabel was a hardcoded English "Scroll to top" while the bottom FAB directly below it already used t('action.scrollToBottom').
The same gap existed in ConversationList ("Scroll to top", "Scroll to bottom") and TerminalOutput ("Jump to top", "Jump to bottom"), five hardcoded strings in total.
New nav.scrollToTop and nav.scrollToBottom keys land in common.json for en, he, ru, and ar, since all three components already load that namespace.
The existing conversation.json action.scrollToBottom key stays in use by the bottom FAB, so no key is orphaned.
The terminal labels change wording from "Jump" to "Scroll" so the announced label is consistent app-wide, and TerminalOutput.test.tsx is updated to match.

Verification

e2e/07_conversation_scroll_gaps.yaml gains a closing block that scrolls back down past the 200pt threshold, waits for the FAB by testID, taps it, and asserts SCROLL-GAPS-TOP-MARKER (the first message) is visible.
That proves the control is still hittable and still scrolls to the top after the restyle, which is a real risk now that the pill is always mounted: an opacity-0 overlay stuck on pointerEvents: auto would swallow list touches, and one that never regains auto would stop working.
All of the flow's existing assertions are intact, including the SPEC-MARKER duplicate-append guard from #470.
The flow already runs in test:e2e:mock, so no script wiring changed.

The full flow passes end to end on a Release build.

• npm run lint — 0 errors, 5 warnings, all pre-existing and none in the touched files.
• Conversation and terminal Jest suites — 26 suites, 277 tests, all passing.
• npm run test:i18n — 3 suites, 55 passed, 1 skipped.

Two pre-existing conditions on this base, neither introduced here.
npx tsc --noEmit reports 14 errors, and the count is identical on the base branch with these changes stashed.
npm run lint:i18n fails its --max-warnings=0 gate on a react-hooks/exhaustive-deps warning in app/session/[id].tsx, and it fails identically on the base branch.

Follow-ups, not in this PR

All three copies of this FAB are now near-identical and could be extracted into one shared component.
Three near-identical copies is a pre-existing condition, so that is deliberately left alone here.

LiveConversationView renders its own inline FlashList and was noted in #470 as deserving the same list-prop treatment; it has no top FAB and is unaffected by this change.