[Task] UI rewrite v2 PR0: perf-gated regression CI

[Astro-Han]

Goal

Automated perf regression gate is active on every PR before any visual rewrite ships.

Scope

In scope:

• Playwright-based perf probe.
• The 4 initial scenarios: homepage-cold, session-streaming-long, tool-call-expand, session-scroll-reading.
• JSON output in the schema described below.
• Base/head comparison rules, ratchet rules, and CI artifact capture.
• Trace-on-failure support as part of the same measurement pipeline when needed by PR0.2.

Out of scope:

• Terminal / ghostty scenario (deferred to PR0.3).
• Lighthouse as the primary measurement path.
• Any visual rewrite itself.

Relevant files or context

Parent: #599

Likely files and existing probes:

• packages/app Playwright suite
• debug-bar.tsx
• renderer-diagnostics.ts

Reference standards:

• INP
• TBT
• CLS
• FCP
• LCP

Key measurement design:

• Playwright Test + browser Performance APIs + same-runner base/head comparison + trace-on-failure.
• Hard fail v1 is regression delta plus catastrophic absolute only.
• Web Vitals good lines stay warn-only until the ratchet promotes them.

Verification

• PR0.1 lands and dev-tip produces at least 3 baseline JSON runs across all 4 scenarios.
• PR0.2 lands and a deliberate regressing PR is blocked by the comparator.
• JSON output includes median + worst values per scenario.
• CI uploads artifacts rather than silently passing.

Execution mode

Agent should investigate and propose a plan first