Goal

The session-scroll-reading perf scenario at packages/app/e2e/perf/perf-probe.spec.ts:470-509 actually exercises the user-visible "scroll stutter" experience and can serve as the regression gate for the core scrolling path tracked under #615. Today it does not — the seeded content is too small, the scroll distance covers only half the timeline, and the input cadence does not resemble real trackpad usage.

Scope

In scope:

• Expand session-scroll-reading to a full-distance round trip (top → bottom → top, covering ≥95% of timeline length).
• Add trackpad-cadence emulation: a 60Hz small-step wheel stream, distinct from the current single large-jump wheel.
• Seed heterogeneous content (tool rows with open/closed states, code blocks, reasoning blocks, long user messages) instead of uniform text.
• Add a long-session variant (100 / 200 / 500 messages) measured separately.
• Add a sustained-scroll variant: 5+ seconds of continuous wheel events, sampling main-thread long-frame distribution.
• Optional: a concurrent-load variant that scrolls while shimmer animation is running.

Out of scope:

• Fixing the underlying perf root causes — virtualization, createMemo overuse, ScrollView throttling, content-visibility / contain-intrinsic-size sizing, scroll state-machine depth. Those are tracked in [Bug] Session UI typing lag and scroll stuttering #615.
• Changing user-visible UI behavior.
• Touching unrelated perf scenarios.

Relevant files or context

Likely files:

• packages/app/e2e/perf/perf-probe.spec.ts (scenario implementation)
• packages/app/e2e/perf/profiles.ts (profile gating / threshold registration)
• packages/opencode/test/lib/llm-server.ts (heterogeneous fixture support if needed)

Related issues / PRs:

• Root-cause issue: [Bug] Session UI typing lag and scroll stuttering #615 (Session UI typing lag and scroll stuttering)
• PR perf(ui): align shimmer duration to 1800ms, add concurrent-shimmer guard, relax perf threshold #631 added concurrent-shimmer-extreme for render pressure on the message-body path. This issue covers the scroll path; the two are complementary, not duplicates.
• PR test: harden session scroll perf guard #635 currently in flight against this issue.

Background signal: hand-test feedback on PR #631 surfaced INP ~200ms on trackpad scrolling that the current session-scroll-reading scenario does not trigger.

Verification

• New scenarios run stably on dev and reliably reproduce the visible stutter / INP regression before any [Bug] Session UI typing lag and scroll stuttering #615 fix lands.
• After [Bug] Session UI typing lag and scroll stuttering #615 fixes land, the same scenarios show meaningfully lower INP and total long-frame time without any threshold relaxation.
• No regression on the existing session-scroll-reading baseline numbers for unrelated probes.
• bun --cwd packages/app test:e2e:perf -- session-scroll-reading passes locally on the low-end profile.

Execution mode

Agent should investigate and propose a plan first.