🤖 perf: measure Perf Profiles milestones in-page; accept the 09-21 workspace-open-large step

[ThomasK33]

Summary

The nightly Perf Profiles workspace-open-large numbers stepped up on 09-21: script ~354 → ~500 ms, task minus DevTools ~469 → ~650 ms, wall ~720 → ~1320 ms. The step is mostly a change in what the spec measures, not extra work. Accept it as the new baseline, and make the harness robust to this kind of change.

What happened

#4293 (4e258e6) changed data-loaded to wait until the tail-first reveal has mounted every row (ChatPane.tsx, the data-loaded attribute). perf.workspaceOpen.spec.ts stops the wall clock and reads Performance.getMetrics when Playwright's toHaveAttribute("data-loaded", "true") passes. Three effects follow:

1. Poll quantization. Playwright polls once, then at +100/+250/+500/+1000 ms (playwright-core server/frames.js). The gap between the in-page data-loaded flip and the passing poll went from a median of 27 ms to 423 ms (local, n=6 each).
2. Window shift. About 100 ms of post-load work used to run after the metrics were read. It now runs inside the window. With a fixed 2 s window after load, the script difference shrinks to +24/+30 ms (two rounds, n=6 each). CPU-profile non-idle time is ~800 ms at both endpoints.
3. Real, intended trade. Full reveal takes +~150 ms (MutationObserver flip time: ~1020 → ~1190 ms). In exchange, the longest main-thread task drops from ~395 to ~161 ms.

The small real leftover (redundant MessageRenderer re-renders on each reveal step, from #4292's inline handleEditUserMessage) is fixed by #4427.

Proposed follow-up (harness only)

• Record milestone timestamps in-page (a MutationObserver on the settled marker, measured from the click). Do not rely on the Playwright poll that happens to observe it.
• Report "useful content ready" (first transcript rows) separately from "fully revealed". Keep longest-task as its own metric.
• Version the measurement contract. When a completion milestone changes meaning, reset the trend baseline instead of flagging a regression.

Do not use a fixed sleep or a vague "quiescence" wait. It was useful as a diagnostic here, but it depends on the workload and adds suite time.

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high

[linear-code]

MUX-196

[ThomasK33]

Done in #4456 (merged as 7b215e4). The workspace-open perf scenarios now record in-page milestones in perf-summary.json: first transcript row, data-loaded (full reveal), and the longest main-thread task. They're timed by a MutationObserver and a long-task observer, not by Playwright's assertion polling. Locally, the polling overhead the old wall number included was 12–195 ms per scenario.

The workspace-open-large step is accepted as the new baseline, with #4427 having fixed the small real leftover. The remaining proposal here, versioning the measurement contract so a trend baseline resets when a milestone changes meaning, belongs with the deferred trend reporter and is tracked in #4442.

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high