1.0 Benchmark Report

Run the public benchmark report before publishing markstream-vue@1.0:

pnpm benchmark:1.0

The command builds the playground, serves it with vite preview, then runs the playground checks that are already used for performance regression coverage:

• Diagnostic Studio baseline, thinking, diff, and stress samples in MarkdownCodeBlock and Monaco modes using /test?benchmark=1, which disables the version sandbox iframe and annotation layer so frame metrics and renderer DOM metrics track the renderer surface.
• Main playground reverse-flex chat initial load, full-scroll pass, and streaming replay using /?benchmark=1.
• Real-browser Web Vitals probes for a million-character restore and scripted scroll path, plus a code-block scroll-into-Monaco path.

It writes:

benchmark/
  results/
    diagnostic-baseline.json
    ...
  1.0.0.chrome-linux-x64.json
  1.0.0.chrome-linux-x64.md
  latest-summary.md

The Markdown summary includes the release package version, git SHA, workspace package versions, Node, OS, CPU, browser, default viewport, row-specific viewport, server mode, LCP, CLS, settle time, frame sample count, frame p95 requestAnimationFrame interval, full-scroll heavy-settle frame p95, max long task, page DOM node count, renderer DOM node count, visible fallback count, heavy-block readiness, scroll drift, Web Vitals restore/scroll phase metrics, and best-effort Chrome-only heap after renderer unmount plus GC when the browser exposes that value.

Initial rows report heavy-block readiness only for blocks visible in the phase viewport, and show N/A when that viewport contains no heavy blocks. Some scenarios use viewport overrides; use the Results table viewport column when citing LCP, CLS, visible readiness, or DOM-node counts. Full-scroll rows report all heavy blocks after the scroll pass. Page DOM nodes are recorded for diagnostics; renderer DOM nodes are scoped to .preview-surface or .chatbot-messages and are the DOM budget used by the release gate. The report also records parse commits, smooth-stream coalesced parse updates, and markdown-it stream parser full/append/tail/cache counters so stream parser regressions are visible in benchmark output. Frame p95 is recorded for every phase; on full-scroll rows, scrollFrameP95Ms covers only the active scroll loop. heavySettleFrameP95Ms records post-scroll heavy-block settle separately. The hard release gate covers fallback counts, heavy-block readiness, DOM budget, CLS, long tasks, settle time, and Web Vitals frame p95/max plus sample-density budgets.

When a scenario fails, the runner continues through the remaining scenarios, writes the full report with passed and failed entries, then exits with a non-zero status.

For script debugging only, set MARKSTREAM_BENCHMARK_SKIP_BUILD=1 to reuse an existing playground build. Do not use that shortcut for release evidence unless the build artifact was just produced. Set MARKSTREAM_BENCHMARK_SAMPLES=baseline,diff only when narrowing a local investigation.

CI workflow

The 1.0 Benchmark GitHub Actions workflow runs on a nightly schedule and can be started manually. It runs the same Web Vitals release-gate mode (MARKSTREAM_RELEASE_GATE=1) used by release:gate:1.0, so budget warnings fail the workflow. It uploads the generated benchmark/ directory as an artifact.

Use workflow artifacts for release notes. Local generated reports are snapshots for their disclosed OS/CPU/browser environment; benchmark/latest-summary.md is a convenience copy of the local report, not the canonical CI latest. These numbers are release regression metrics, not universal performance guarantees. Do not claim speedups that are not present in a generated report.

Release gate

The 1.0 release gate runs:

pnpm run release:gate:1.0

That command executes release:verify, docs:build:ci, size:check, and the benchmark report. The release command uses .tmp/benchmark for the generated report so publishing does not depend on tracked benchmark artifacts.

Before publishing, run the full package dry run:

pnpm run release:dry-run:1.0

That repeats the release gate, then dry-runs the parser, core, and Vue package publish path.