test: add margin to warm-switch performance gate

[wesbillman]

Outcome

Keep the warm-switch <100ms target, with an owner-requested provisional <200ms per-sample hard ceiling. Record all four timings and finish the existing correctness checks before enforcing the ceiling. Target misses produce performance annotations in the downloadable Playwright JSON; raw frame/first-visible diagnostics remain in evidence.json.

No production changes, retries, extra samples, timing-boundary changes, timeout increases or browser-selection changes. The 1s completion watchdog, correct-channel row/composer checks, held-label cold-open proof, no-extra-head-read check and profile/relay assertions remain intact. Existing Linux WebKit exclusion remains historical/awaiting revalidation, not an active 100ms gate.

Why this policy

The same single-sample assertion failed on main at 118ms, #182 at 111ms, and #185 at 120.5ms. These are intermittent observations, not proof of contention or absence of PR-specific cost. 200ms is an explicit margin choice with roughly 80ms above the largest observed miss, not a proven flake-free bound. 100–199ms regressions now need performance review; >=200ms still fails. Severe runner stalls can still fail.

Validation

• Clean commit 96338eadd06b9cf295b1ea99ed4fe999ec0aba61: complete channel-opening.spec.mjs passed 4/4 across serial Chromium 153.0.8010.12 and WebKit 26.6 on macOS ARM64, Node 24.18.0. Command: bin/pnpm exec playwright test --config tests/browser/playwright.config.mjs tests/browser/channel-opening.spec.mjs --project '*-measurements' --no-deps --workers=1 --reporter=list,json.
• Transient local copies of the full test file (not shipped), both engines: adding 125ms inside the measured interval failed the old gate at 164/184ms, but passed the new gate at 149–189ms with four annotations and complete relay/profile checks. Adding 250ms failed every measured 275–303ms switch at the ceiling, after all four samples and functional checks. Suppressing warm navigation failed the 1s watchdog. An actual additional fixture HTTP head query failed the request-count assertion. These probes validate gate behavior, not real runner contention or hosted stability.
• Independent source review found no blockers; documentation clarified annotations are in the report artifact, not the GitHub job-summary table. Mandatory commit/push hooks passed; their selectors correctly skipped unrelated unit/design work. No hooks bypassed. No new or removed browser cases; the existing browser-only layout/paint and production-broker contract is retained.

Local cost comparison

• fe897a10: 4 passed; elapsed 14.31s, summed execution 10.19s; compiled setup Chromium 647ms / WebKit 696ms.
• 96338ead: 4 passed; elapsed 11.84s, summed execution 9.45s; compiled setup Chromium 580ms / WebKit 614ms.

One local run per snapshot is a cost sanity check, not a distribution or Linux result. Hosted full-suite CI and reviewer/code-owner approval remain pending. This PR does not approve #185 or automatically merge anything.

Implemented by Carl, an automated agent, via Wes’s GitHub account. Source conversation: Buzz channel 3428ec3f-a58b-429b-afbc-dc6a77918ce8, thread bf84443949508192684e6cb0049d775dea51ef08d82c6f93d122c639330d0251.