fix(conversation): stop scroll-up blank gaps and anchor teleports

[RonenMars]

Problem

Scrolling up through a paused conversation opened progressively larger blank regions where messages should be, and a small upward scroll could teleport the viewport far back in the conversation.
Space was reserved (the scroll indicator stayed proportional) but rows were not drawn, gaps healed on a downward scroll, and they repeatedly opened directly below a collapsed Reasoning row.

Root causes

Three independent defects compound here.
All three were reproduced and measured against the real session that produced the report (80 messages, paused, live streamer) with temporary instrumentation before any code changed; the instrumentation is not part of this diff.

1. The delta poll grew a paused conversation, one duplicate row per tick

The streamer's after_index window is inclusive — [after_index, after_index + limit) — but the tail view resumed from deriveCursor, which is the max message index the client already holds.
So every poll re-fetched the last message and stored it as a new page, and mergeConversationPages emitted it again with the same ${convId}-${index} id.
Measured on the untouched paused conversation: the rendered list grew from 80 to 156 rows during the capture, roughly one phantom row per 6 seconds, every copy carrying the duplicate id <convId>-80.
Duplicate ids break both FlashList's keyExtractor and the maintainVisibleContentPosition anchor, which stores the first visible item's key; each append also added ~342 pt of content height and triggered an autoscroll plus an anchor pass underneath the user's finger.

2. drawDistance was left at the 250 px default against rows several viewports tall

drawDistance is pixels of pre-rendered runway, not rows, and FlashList v2 splits the 2 × drawDistance buffer 70/30 toward the scroll direction — scrolling up pre-renders only ~350 px above the viewport and evicts rows ~150 px below it.
Measured row heights in this conversation: 3,114 pt, 2,146 pt, and 1,892 pt against a 654 pt viewport, so a single markdown-table answer is roughly five viewports tall.
Any real flick outruns that buffer and the just-passed row unmounts into a visible blank, which is why the gaps opened directly below collapsed Reasoning rows: those are the cheapest rows to draw and the last ones that still made it.

3. One coarse getItemType pool poisoned the estimated-height average

FlashList v2 positions not-yet-measured rows from a per-item-type running average (heightAverageWindow; see Shopify/flash-list#2053 for the mechanism).
A 46 pt collapsed thinking row shared the tool pool with tool and diff cards, and 300-character replies shared assistant with ~3,000 pt table answers, so estimates for unvisited regions were far off and every real measurement re-laid-out everything around it.
Measured content-height thrash between consecutive scroll events: 18k → 45k → 42k → 57k → 66k pt, with a maximum single relayout of ±19,898 pt.
maintainVisibleContentPosition has to absorb that as offset corrections, and the captured pathological case is exactly the reported teleport: a +700 pt offset jump paired with relayouts of +12,174 pt and −3,490 pt.
This is the maintainer-documented failure regime for items taller than 2 × drawDistance (Shopify/flash-list#2136 and its fix PR #2138; the anchor-loss half shipped in 2.3.1, the slow-convergence half is still open upstream as #2052 and #2018).

Fixes

• hooks/useConversations.ts — the { resume } delta now requests after_index = maxHeld + 1. The { after } variant already carries the server's own next_after_index, which is the first index the client does not have, so it is unchanged.
• hooks/useConversations.ts — mergeConversationPages dedups by message_index at the single flatten choke point, so overlapping server windows can never produce duplicate row ids again regardless of transport behavior. This also covers anchored windows, which the server deliberately widens backward near the tail.
• components/conversation/ConversationHistoryList.tsx — drawDistance={2000}, which keeps the engaged window ahead of momentum scrolling and clears the item > 2 × drawDistance regime for rows up to 4,000 px (tallest measured: 3,114 px).
• components/conversation/ConversationHistoryList.tsx — getItemType split into thinking / diff / tool / user / assistant, so each recycling pool's height average tracks rows that actually look alike.
• components/conversation/ConversationHistoryList.tsx — ListHeaderComponent and ListFooterComponent element identity is memoized. Inline conditionals recreated them on every parent render, and upstream guidance (FlashList v2 Chat UX Regression: Inconsistent scroll behavior with paginated messages Shopify/flash-list#1844) flags that as an anchor-churn source during onStartReached fetches.
• components/conversation/ConversationHistoryList.tsx — the animate-id bookkeeping walk is keyed on the messages array identity instead of running in every render body.

Verification

Same scripted scroll protocol, before and after, on the real session against the live streamer.

Metric	Before	After
Data churn while paused	78 appends, 80 → 156 rows, duplicate ids every tick	0 appends, no duplicate ids
Content-height relayouts > 400 pt	19, max ±19,898 pt	4, max 1,742 pt
Offset jumps > 600 pt	3, diverging from the paired content shift (+700 pt against +12,174 / −3,490 pt)	4, each matching its paired relayout within ±40 pt
Fast-fling gauntlet to the top	blank viewport frames	every frame dense, lands on the first message
Rows whose height changed after first layout	0	0

Ruled out

• Re-render storm from the FAB setState at scrollEventThrottle={16} — component re-renders during scroll measured at approximately 0 per second, since React bails out on same-value sets. Cell renders only spike during mount fill. The scroll handler is unchanged.
• Entrance animations in recycled cells — zero rows rendered with animateIn=true across every paused-session capture.
• Collapsed Reasoning rows reporting expanded height — ThinkingCard, ToolCard, and DiffViewer all render collapsed content only via useRecyclingState(false), and their measured heights are honest. The gaps neighbor these rows because they are the cheapest rows to draw, not because they misreport.
• Post-mount height settling from markdown or prism — zero rows changed height after first layout in either run. Tables render as wrapped text and prism tokenizes synchronously.
• onStartReached re-fire or pagination churn — zero startReached fires during the capture, since the conversation fits in one 80-message page. The header-identity fix hardens this path regardless.
• Offset projection over-projecting under slow renders — averageRenderTime is clamped to the [16, 32] ms range in RenderTimeTracker, which bounds projection to roughly 100–160 pt per event, far too small to produce the observed displacement.

Tests

__tests__/unit/hooks/useConversations.test.tsx — the existing resume-delta test pinned the off-by-one by asserting after_index=2 for a held max index of 2; it now asserts after_index=3 and rejects after_index=2.
A new test feeds an overlapping delta page and requires the merge to produce unique ids.
On the pre-fix code these two fail and the other 53 in the file pass.

e2e/07_conversation_scroll_gaps.yaml with e2e/fixtures/conv-scroll-gaps.json — the fixture's height distribution mirrors the measured session: two roughly 3,000 pt markdown-table answers and a tall code fence among 50–350 pt rows, with collapsed thinking rows at the gap edges.
The flow dwells at the tail so the delta poll runs, asserts the tail has not grown, then flings to the top with no settle time and requires the top rows drawn within a tight timeout.
Verified failing on a pre-fix Release build (at the tail-growth assertion) and passing on the fixed Release build.
Wired into test:e2e:mock.

The flow pairs inline rather than via setup.yaml because Maestro 2.6.1's hideKeyboard does not dismiss the iOS 26 keyboard; the prelude scrolls the form instead and submits with the token field's return key.

Notes

__tests__/e2e/onboarding-flow.test.tsx has one failing test on this branch, and it fails identically on the base branch with these changes stashed, so it is pre-existing and unrelated.

LiveConversationView renders its own inline FlashList with the same default drawDistance and no getItemType. It benefits from the data-layer fixes here, but its list props deserve the same treatment in a follow-up.

The ↑ Top FAB overlaps list content while scrolling, which is visible in the original report's screenshots. Separate cosmetic issue, untouched here.