fix(site): split Working blocks at assistant text and scroll to their end on expand - #30656

Deliberate choices from self-review (frontend-review FE1 to FE10 and coder-frontend-review):

• behavior: "smooth" bypasses the reduced-motion rule in site/src/index.css (scroll-behavior: auto), as the existing chat scroll commands already do (submitChatTurn.ts, ConversationTimeline.tsx).
• The last-row reveal jumps instead of gliding: following pins the end with an instant scroll on the next resize, which cut a smooth scroll short after one frame (measured in Chromium).
• Following a growing block starts only after a reveal that scrolled, and ends 2 seconds after its last scroll starts, so later growth (a re-render, a late image) does not move a reader who stopped scrolling. A smooth glide uses part of that window. Each chase scroll renews the window, so images that load one after another stay followed. A width change ends it at once: a resize re-wraps the steps taller, and that is not growth the reader asked to follow.
• Following a growing block stops on any scroll up, including one with no gesture (a focus move, find in page): the reveal only scrolls down, so a scroll up came from somewhere else. It only scrolls when growth pushes the end below the view, so it never pulls the view back up.
• A block taller than the view scrolls its header above the top to show its end. Keyboard focus stays on the header button.
• If rows below an expanded block, such as its attached files, already used up the anchor hold's tail spacer, the hold stays and the view does not follow the reply, as on main. Remote UAT hit this at 1440x560; at 1440x700 the view handed off. Handing off in that case would mean jumping to the end at the expand, which moves the block the reader just opened.
• Collapsing a block that was revealed at the end of the chat can drop its header down the view: the content shrinks and the browser clamps scrollTop. main jumps the same way from the same scroll position; the collapse path is unchanged here.
• The harness stubs requestAnimationFrame with setTimeout. jsdom runs frames on the global setInterval, and the existing live-label test fakes timers while a motion component requests a frame, which stalls every later frame in the file.
• The test-only BlockRow stays separate from ToolCall.test.tsx's ToolRow: sharing them would add a render-prop abstraction for two callers.

Working block grouping:

• Files attached in the middle of a turn no longer split a block. They render after the block's last step, so a block that attaches several files shows all of them below it.
• File-reference parts stop closing blocks too. Only user messages carry them, and a user message already closes the block.
• LabeledNarrationFolds and LeadInBeforeSearchFolds are now LabeledNarrationStaysVisible and LeadInBeforeSearchStaysVisible, so Pixel sees them as new stories.
• frontend-review FE9 asks for story variants in named constants, while coder-frontend-review TS3 and TS4 ask for overrides at the call site and no single-use story constants, and that skill lets its checkpoints win. The two new single-use message variants in ConversationTimelineWorkingBlocks.stories.tsx stay inline, like the file's existing ones.
• The mock attach_file result carries only name, the one field AttachFileLabel reads.
• Reasoning at the start of a turn, before the first text, renders on its own above that text: the text closes the block, and completed reasoning with no steps stays visible, as main already shows for a turn with no tools. While that reasoning streams, a live Working block shows and dissolves when the text arrives.
• When one render brings a live-only block's prompt page, its persisted step, and a cleared stream, the block keeps its scroller key by matching the start it recorded while live. The stream carries no reasoning timestamps, so a block that starts after text and streams reasoning before its first tool call can still remount in that render.

Row spacing:

• The 12px gap applies to every transcript row, so Pixel shows a diff on every story that renders a chat transcript through ChatMessageScroller (the ConversationTimeline, ChatPageContent, AgentChatPageView, and AgentChatPage stories). The WorkingBlocks and LiveWorkingBlocks stories render a bare MessageScroller.Content with no gap, so they do not change.
• The vendored scroller reads the computed row gap (getFlexGap in geometry.ts), so its spacer math follows the class.

• Remote dogfood UAT through Coder Agents on an isolated dev deployment, desktop viewport, 8 scenarios on ccbb338f5cb. Round 1: FAIL on the collapse clamp above, which reproduces on main from the same scroll position; that scenario was reworded to record the clamp as an observation. Round 2 on the same SHA: PASS. That run did not test a turn split across history pages or non-desktop viewports; the later run below does. Between runs, unit tests and a Chromium probe story (not committed) that records every scrollTo call and the scroll position each frame cover the changes.
• Nine new WorkingBlockDisclosure scroller tests (11 cases). Removing the reveal call, the prop, the !expanded guard, the in-view check, the anchored hand-off, the whole anchored branch, the spacer check, the growth chase, its settling-jump check, its in-view check, either height record, the clear on another jump, or the clear on a scroll up each fails at least one of them. The chase's growth check is only visible in a browser, where jsdom's instant scrollTo hides it: in the Chromium probe, removing it re-issued the smooth scroll on every resize during the glide (3 calls instead of 1 when a row below the block grew mid-glide).
• Chromium probe on the final code: a block that grew 300 px mid-glide got a second smooth scroll at the growth and settled with its end 24 px above the bottom of the view; growth after the glide settled did the same; after focus() moved the view up to an earlier row, a later growth left the view there.
• Codex round 1 on 715f799a8ed: two P2 findings (a cut-short smooth reveal at the last row; no hand-off when the expansion used up the anchor spacer). Both confirmed by execution and fixed in c7f98a67d99.
• Codex round 2 on c7f98a67d99: one P2 finding (content that grows after the reveal, such as a lazy image, pushed the end back out of view). Confirmed with a failing test and fixed in 990cb3542af. Self-review then found that a scroll with no gesture left that chase armed; fixed in b928354858d and d7053a2de50.
• Codex round 3 on d7053a2de50: no findings.
• Rebased onto main at 99afdc3b0fb to pick up fix: fold web searches, answer reasoning, and OpenAI narration into working blocks #30529, which folds an answer's reasoning into the Working block it ends. The first commit now also exports isViewport from testHelpers/messageScrollerLayout.tsx, because the ToolCall.Scrollable tests from refactor(site/src/pages/AgentsPage/components/ChatElements/tools): add ToolCall.Scrollable #30648 use it and this PR moved it there. Commit SHAs above are from before the rebase; each rebased commit keeps its subject. Codex round 4 on the pre-rebase merge bde29b9a1cc: no findings. Codex round 5 on ab2fdeec307: no findings.
• Self-review: coder-frontend-review (eight roles) and frontend-review (FE1 to FE10) over five rounds, the last on the Codex fixes, re-running each failing role after its fix until every role passed on d7053a2de50; findings fixed or justified above.
• Working block grouping, f97f2a911d3 to 5f87b148d88: written and reviewed on a separate branch, then cherry-picked onto ab2fdeec307 with no conflicts and the same patch (git patch-id). coder-frontend-review ran 13 rounds until all eight roles passed, and frontend-review FE1 to FE10 passed apart from the FE9 note above. Red-green: letting attached files close the block fails both new attached-file tests, and the old live rule fails the live one.
• On 5f87b148d88: tsc, biome, oxlint (including the design-system config), knip, the React Compiler check, and dpdm pass, and so do the unit tests in src/pages/AgentsPage, src/vendor, and src/testHelpers (2128 passed, 2 skipped) and the ChatConversation story tests (118).
• A long dogfood chat with attached files, loaded read-only and rendered with the attached-file fix: the reasoning before the final answer is the last row inside its block, and the files and the answer follow the block.
• Row gap, aac53217c4a: compared 8px, 12px, and 16px on the same read-only dogfood chat in Chromium at 1280 and 390 px widths. The story tests that render ChatMessageScroller pass (7 files, 215 tests), and coder-frontend-review (eight roles) and frontend-review FE1 to FE10 pass.
• Review by @jaaydenh on aac53217c4a, fixed in 29a3bcb14b6: the growth chase had no end besides user input, and an expand forced up to four full transcript layouts. The chase now expires and ends on a width change. The reveal runs before the anchor hold, which it replaces when it scrolls, and the row's rect is read once, before the scroll's writes. Two new cases (the window passes, the block re-wraps at a new width) each fail with their check removed. The reorder changes no test outcome (29 pass either way); in a throwaway jsdom probe, an anchored expand made 21 getBoundingClientRect calls instead of 69. Not measured in a browser. coder-frontend-review passed after four rounds; the size check reuses SCROLL_POSITION_EPSILON, which upstream also applies to row positions. Site checks pass on the new head (unit 2130 passed, 2 skipped; ChatConversation stories 118).
• Codex round 6 on 29a3bcb14b6: two P2 findings. When the newest loaded page opened with files the agent attached and later rows continued their block, the files rendered above a block labeled complete, and loading the older page moved the block above the files. Confirmed with a failing test and fixed in ffcae5a6b7e: the block on that page is partial and renders above the files from the start (story PaginatedFromAttachedFile). Declined: running the scroller tests in a real layout engine. They assert the rows the scroller publishes as visible, with the jsdom model main already uses, and the Chromium probe and remote UAT cover browser behavior.
• Codex round 7 on ffcae5a6b7e: five style and structure findings. Fixed in acbd47dcc88: a hidden-entry assignment that nested a ternary, a nullish fallback, and a conditional, and a doc comment that restated a field name. Declined: extracting the row renderer into a component (the closure is on main, and the file goes from 795 to 798 lines), a named props type for the Transcript test helper (moved unchanged from ToolCall.test.tsx), and splitting a streamed row once rather than once per slice (the live row holds only the current step; in jsdom, a streamed update cost 0.04 ms for 3 sections and 1.8 ms for 50).
• Codex round 8 on acbd47dcc88: one P2 finding. A live-only block that starts after text in the same stream remounted, collapsed, and lost focus when its prompt page, its persisted step, and a cleared stream arrived in one render. Confirmed with a failing test and fixed in 8bb9e94b76b. Codex round 9 on da079c0b460: no findings.
• Remote UAT through Coder Agents on dogfood: an isolated dev deployment, real Anthropic and OpenAI models, desktop and 390 px viewports. Run 1 on 29a3bcb14b6: the reveal, the last row, the anchored turn, growth follow, grouping with real models, the row gap, collapse, and the mobile viewport pass, and the collapse clamp matches main. It also found gaps that main does not handle either: an anchored expand of a block with attached files below it never handed off to following (da079c0b460 fixes it while some tail spacer is left), an expand within about 30 ms of an older page starting to load loses the reveal, and growth after an expand that did not scroll is not followed. Part 2 on acbd47dcc88: with the page boundary at the attached files, the partial block renders above the files before and after the older page loads (the control, 29a3bcb14b6, reorders them); the reveal, growth follow, and real-model grouping pass, and the grouping refactor renders 4 web-search chats the same as ffcae5a6b7e. Part 3 on da079c0b460, with acbd47dcc88 as the control: an anchored expand of the live block with attached files below it hands off to following at 1440x700 (2 of 2 runs, control 0 of 1), and so does an expand of an earlier block of the turn (2 of 2, control 0 of 2); the reveal below the fold and the anchored expand of the last block pass. At 1440x560 the files used up the tail spacer before the expand, so the hold stayed, as on the control and main (see the review notes). Verdict on da079c0b460: PASS. Not tested: an OpenAI lead-in before a web search, OpenAI narration across three or more tool calls, and a real image that loads late inside a block (the growth checks injected it).
• On da079c0b460: tsc, biome, oxlint (including the design-system config), knip, the React Compiler check, and dpdm pass, and so do the unit tests in src/pages/AgentsPage, src/vendor, and src/testHelpers (2134 passed, 2 skipped) and the ChatConversation story tests (119). coder-frontend-review (eight roles) and frontend-review FE1 to FE10 pass on each fix, with the failing roles re-run after their fixes.
• CI on da079c0b460: the required checks pass. The Storybook (Pixel) job fails on one story this PR does not touch: the UserEngagementChart Loaded story's play function cannot find its "User engagement" group. fix(site/src): replace role="application" on recharts charts #30375 added that check to main with the wrong case, and fix: match chart accessible name in story #30907 fixed it on main after this run started. A re-run reuses this run's merge commit, so the job stays red until the next push.