perf(web): stop Composer and the chat header re-rendering on every stream event

[Andrey Markin (Mark-Life)]

Stack (4/5): #616 profiling harness → #617 heavy-load scenarios → #618 streamed markdown blocks → this PR → #626 Shiki worker.

Depends on #618. This PR targets main, so it also shows the commits of #616–#618. Review only the last commit: fbff8755.

Problem

Composer and the chat header re-render on every stream event, even though nothing they show changes. During a 26k-char streamed answer, Composer renders 175 times. The closed plan popover in the header re-renders its whole Radix subtree, more than 5,000 component renders.

Approach

Two causes, each fixed at the source:

• Composer: ChatView rebuilt recentPrompts from turns on every event, so Composer got a new array each time. ChatView now keeps the previous array while its contents are equal.
• Chat header: useChatTodos wrote a ref during render, so the React Compiler skipped the hook. It returned a new plan object on every render, which re-rendered ChatHeader and the plan popover. The ref now syncs in an effect, and a try/catch around a value became .catch(), so the hook compiles.

Radix does not mount closed popover content. A closed popover costs only when its parent re-renders, so fixing the parent is enough.

Changes

• chat/ChatView.tsx: stable recentPrompts.
• chat/useChatTodos.ts: ref synced in useInsertionEffect; reloadPlan uses .catch().
• apps/web/SPEC.md: useChatTodos leaves the list of compiler bail-outs (62 → 61).

Screenshots

Same 26k-char stream, CPU 4x, react-scan outlines, mid-stream. Before: Composer ×102, the send-button Popper ×103, and the plan popover at the top. After: none of them re-render.

Before	After

Checklist

• Fast gates pass: bun run lint, bun run typecheck (17/17), bun run test (web 1,411 pass). compiler:census: 408 compiled, 61 bail-outs, no new ones
• E2E suite passes for app-affecting changes (bun run e2e on this branch, top of the stack): 3 failures, none from this stack. 00-jbcentral.spec.ts:20 fails the same way on main. activity-breadcrumbs.spec.ts:240 fails 1 in 12 runs on both main and this branch. layout.spec.ts:1061 passed 3/3 when re-run alone
• Before/after screenshots are included for frontend changes, or marked not applicable
• Relevant SPEC.md / top-level specs updated to reflect any boundary, contract, or behavior change
• I have read the Contributing guide and agree to the Code of Conduct

bun run perf:render --runs 1 --cpu 4, 3 interleaved rounds, median [min–max], base #618 vs this branch:

scenario	Composer renders	popover subtree renders	self ms of these	total render ms
chat-streaming	268 → 2	7711 → 73	192 → 4	1535 → 1555
long-stream	175 → 2	5167 → 23	155 → 3	1459 → 1435

Render counts and self time drop well outside the spread. Total render ms does not change beyond noise: commits get slightly cheaper, and React batches fewer of them. Measured before the stack was rebased onto current main; the change it measures is identical.

[jetbrains-air]

Approved — ready to merge.

[jetbrains-air]

Approved — ready to merge.