Production resource budgets

e2e/performance.spec.ts loads cold production pages and counts first-party JavaScript, CSS, fetch, and XHR response bodies through browser-context events. This includes worker requests; it does not rely on window resource timing. HTML bytes and inline style bytes are measured separately. The tests use the same Chromium production preview as the navigation regressions.

The gate allows 10% growth over the measured baseline below. It protects cold visits and first search use as the corpus grows. It is a regression budget, not a device-independent LCP/INP target or a compressed transfer measurement. Images, fonts, PDF content, and third-party integrations are outside this code budget. Tests stub external requests; rendering.spec.ts complements those stubs with the actual pinned MathJax engine and fonts.

Measurement: 2026-09-27, 105 articles, Node 24.21.0, Angular 22.2.0, and Playwright 1.63.0 Chromium on the served production build. The old 2026-09-12 and 2026-09-20 limits predated later article and search-corpus growth. On an isolated checkout of the unchanged commit with Node 24.18.1 and the previous lockfile, the same three code and search gates failed: home 422,741 B, blog 692,172 B, CF551C 799,490 B, and first search 2,194,246 B. The dependency refresh measured 426,564 B, 695,835 B, 803,370 B, and 2,194,165 B respectively. The new baselines retain 10% headroom and track the current corpus and build output. Static formula CSS remains included in HTML, not an additional transfer.

A local sample recorded 83 ms to focus the search input, 218 ms from opening to the first Chinese results, and 20 ms for a warm English query. These host timings are diagnostic observations and are not asserted. The search behavior suite instead holds the engine download open and verifies that typing, dismissal, and focus recovery already work. Warm queries must request no additional code.

Each budget test writes resources.json with the exact URLs, sizes, and search phases, and attaches it to the Playwright report. When a budget fails, inspect those requests before changing the baseline. Explain intentional growth in the change description; do not raise all budgets because one article grows. Keep unrelated diagram styles out of this representative article.

The report also includes a separate delivery list for all observed first-party responses, including documents, fonts, images and worker requests. It records Content-Encoding, encoded response-body bytes from Playwright request sizes, and decoded body bytes, split between initial reading and cold search. These exclude HTTP headers but may include transfer framing reported by the browser. The Node preview normally uses identity; its results must not be labeled production compression savings. Lazy images not requested by that journey and stubbed third-party resources are outside this observation. Existing decoded-code budgets remain independent of these delivery diagnostics.

Long article runtime

Payload budgets do not measure the browser's formula-processing or scrolling cost. When reviewing math-heavy articles, also use the real pinned MathJax engine and distinguish startup tasks, settled scrolling, and work deferred until a section becomes visible. An engine download being lazy does not mean its per-article typesetting is lazy. Closed disclosures can still incur that work.

Real-engine rendering tests check correctness; they do not currently enforce long-article runtime budgets. Record runtime measurements with their tested revision and conditions before using them to justify an optimization.

Generated articles now ship complete CHTML formulas and local fonts. Reading, including opening disclosures, selecting, copying, or right-clicking formulas does not load the browser MathJax engine. Angular-authored views with raw TeX still use the on-demand runtime engine, with its custom context menu disabled. Initial navigation adopts prerendered HTML, whereas subsequent SPA navigation loads a single article module. Assess both paths: larger HTML is a deliberate tradeoff for removing client typesetting, not a claim that raw markup shrank.

After pnpm build, run pnpm exec playwright test e2e/long-article.spec.ts --project=desktop for RL, Diffusion and SSL at a 390px viewport and Chromium 4x CPU slowdown. Each test attaches startup/scroll long tasks and heading geometry reads, then repeats with a test-only doubled article DOM. The 90 frame-stepped scroll jumps are a stress diagnostic, not a real-user FPS benchmark. Timings are reported rather than gated; body duplication and unsolicited engine loads are gated. Use the static and dynamic formula rendering suites alongside this diagnostic.