Session-export load benchmark

Replay an exported session against a production app build. The two cases compare the default Compact preset with every category ungrouped and details still collapsed.

From packages/app in PowerShell:

$env:PLAYWRIGHT_BUILD = '1'
$env:PLAYWRIGHT_BASE_URL = 'http://127.0.0.1:4398' # Existing production preview
$env:LAGGY_SESSION_FILE = 'C:\path\session.json'
$env:LAGGY_SESSION_OUTPUT = 'C:\tmp\opencode\session-load'
$env:LAGGY_SESSION_HISTORY = 'paged' # Or 'full' to supply all exported history
bun x playwright test --config e2e/performance/playwright.config.ts timeline/laggy-session-benchmark.spec.ts --repeat-each=20 --workers=1 --retries=0
bun e2e/performance/timeline/laggy-session-report.ts $env:LAGGY_SESSION_OUTPUT

Each test uses a fresh browser context and measures one cold load, switches back to the source session, then measures one warm load. Repetitions therefore interleave cold → warm pairs rather than collecting separate cold and warm batches. There are no discarded warm-up switches. The warm member of every pair must issue zero message requests. Each pair is saved in a separate JSON file; compare paired differences as well as the cold and warm distributions when system load varies.

The report writes summary.json and prints the median, p95, maximum, and median paired cold-minus-warm difference. It rejects incomplete or cold-only records instead of mixing them into paired results.

--repeat-each=20 collects 20 pairs per grouping mode. LAGGY_SESSION_COLD_ONLY=1 remains available for focused cold profiling. Screenshots are taken after a pair finishes, not between its measurements.

The app shell, source session, model control, and fonts are ready before the timed action. These measurements cover session entry, not application startup. firstCorrectObservedMs begins at mousedown and ends when the destination is visible at its expected bottom position, including Compact's automatic history fill. stableObservedMs includes three-observation confirmation and must not be treated as additional rendering time.

Set OPENCODE_PERFORMANCE_TRACE_DIR for Chrome traces. LAGGY_TRACE_ITERATION=0 traces the cold load; the default (1) traces the warm member of the pair. Profile separately from timing runs.

LAGGY_HTTP=1 disables route interception for an external HTTP replay server containing the same export and source fixture. Keep direct HTTP and Playwright-routed cold series separate: routing adds transport overhead. Raw samples, mode settings, viewport, browser version, and screenshots are retained in the output directory. The export itself is not copied into the repository.