[mfal]

What & why

Turns findings of the perf harness (#3524) into ordinary browser tests. They count, never time, so they hold on shared CI runners and run in the existing browser job.

List (List.renderCost.browser.test.tsx)

• Weight per item: at most 340 fibers / 35 DOM nodes for a rich item (Avatar, Heading + AlertBadge, Text, ContextMenu), and 92 / 13 for a simple one. Today's counts are 336 / 34 and 89 / 12. The file says how to move a ceiling, and a failure prints the new count.
• Render functions: with stable render functions, textValue and onAction, neither a parent re-render nor a sort calls an item render function.
• Item subtrees: a Profiler around each item shows whether its subtree commits, with or without a render-function call.
  • An update next to the list updates no item.
  • A parent re-render and a sort update all 20 items. That is the ListState context cascade from Runtime performance of Flow components: measurement results (2026-10) #3528, so both are test.fails. Drop .fails once the cascade is gone.
• State updates come from outside, not from clicks. A click changes react-aria's global interaction modality, and every item follows it, so a clicked re-render reported 20 items even for state the list never sees.

DOM identity on a re-render with equal props, following the Markdown (3808e79) and CodeEditor (#3467) tests:

• Table: rows and cells
• Select, closed: the trigger and its field
• Markdown: the CodeMirror editor of a code block

countFibers(container) in src/lib/dev/fiberTree.ts is reduced from the harness's fiber walker, because #3524 is not merged.

How the tests were verified

Each guarded property was broken temporarily, and its test failed:

break	fails
Item memo comparator always false	both render-function tests (20 and 40 calls)
+5 pass-through components in Heading	both fiber ceilings (rich 341, simple 94)
+2 spans in Heading	both DOM ceilings
key={Math.random()} on the Table scroll area, Select's Aria.Button, CodeMirror	the matching identity test
Math.random() in the deps of Markdown's renderer memo	the code-block test, plus the existing one
test.fails → test	both, on expected 20 to be +0

WebKit: 3 runs in a row are green. Firefox does not start locally (macOS 27), so the non-blocking Firefox job is the first check of whether react-aria renders a different count there.

Issue

Part of #3528

Checklist

• PR title is a Conventional Commit and matches the base branch above
• The issue this PR finishes (if any) is linked with Closes #… (finishes none)
• pnpm lint is clean and the touched browser tests pass (WebKit)
• Generated code is committed (none affected)
• User-facing strings: none
• Docs: no public API change

🤖 Generated with Claude Code

[github-actions]

Coverage Report for ./packages/components/

Status	Category	Percentage	Covered / Total
🔵	Lines	64.82%	1410 / 2175
🔵	Statements	64.28%	1449 / 2254
🔵	Functions	50.63%	318 / 628
🔵	Branches	58.94%	758 / 1286

File Coverage

No changed files found.

Generated in workflow #7462 for commit 37270a2 by the Vitest Coverage Report Action

[github-actions]

🚀 Preview Deployment

Preview environments are ready:

Type	URL
docs	pr-3569.docs.review.flow-components.de
storybook	pr-3569.storybook.review.flow-components.de

Images:

• docs: ghcr.io/mittwald/flow/docs:pr-3569
• storybook: ghcr.io/mittwald/flow/storybook:pr-3569