Production,
under pressure.

In mobile 04, 45 script requests completed 3.33 MB before DOMContentLoaded at 8.006 s. The three large assets below were parser-initiated. Their rendered script elements later showed neither async nor defer.

These are strong leads for a delivery experiment, not a proven module graph. Bundle names do not identify the expensive package; rendered attributes do not reconstruct every original parser dependency. The LCP image’s initial request priority was Low in all ten runs. Priority changes were not analyzed, so no final-priority claim is made.

Request evidence · Rendered script attributes.

Three additional 15-second checks keep the mobile viewport, user agent and DPR constant. They vary CPU slowdown and network shaping. Each alternative setting has one observation and is not pooled with the five-run baseline.

Removing CPU slowdown while retaining 4G did not remove the long LCP. Native network shortened it substantially even with 4× CPU slowdown. This directionally supports investigating the delivery envelope; it is not a measured optimization, a causal effect estimate or a forecast of real-user gains.

“Catalog anchor” is the first 50 ms-polled geometry match for a catalog link. It is not click readiness, accessible interaction readiness or proof that every card has loaded.

These are accepted-session loads with cookies/storage retained. They do not measure a brand-new user’s legal acceptance, challenge, consent or authentication path.

HTTP totals count request hops and completed encoded bytes within the recorded window. Pending/failed partial transfer is excluded. Raw capture totals also contain data/extension resources; the report filters those out. All cache-served entries in the raw landing captures were non-HTTP data resources. Cross-origin ResourceTiming transfer size of zero is not cache evidence.

Only the parent browser target is observed. Worker and out-of-process iframe traffic may be absent. Landing network counts have a small final-drain boundary gap. Room and diagnostic captures include a drain after the observation boundary. No native HAR, headers, response bodies or physical-wire byte total is claimed.

The slower mobile run remains included because its recorder and source qualification are valid. No performance-based outlier removal is applied.

A frame requires a changed pathname, visible video geometry, and either a newly discovered video element or post-click media source/generation change. Every measured frame used an existing element with a changed source and generation 1. This prevents a landing preview’s old frame from counting as room startup.

Control readiness is the first visible, enabled button or button-role element with one of the documented player aria-labels. It does not verify every player operation or keyboard path. A 50 ms poll bounds these first observations; visibility checks do not establish occlusion-free viewing.

Frame callbacks concern frame submission/composition, not guaranteed physical display. A playing event may arrive after a frame callback, so the report does not force them into a false chronological sequence. Video-frame callback specification.

The query flag is absent from every room URL. Same-document time origin and modern-interface markers persist. That establishes continuity of the observed document, not server-side experiment assignment after navigation.

One mobile journey has 134 carry-in requests and about 1.16 MB of whole-hop carry-in completion. Those bytes can include transfer before the click; they are not additional click-window bytes. Separate ongoing landing work from new room work before deciding what to cancel.

Every included room capture has zero post-click and zero unknown-time orphan events. Remaining orphan events are explicitly before the click. Three recordings without that classification are excluded from the reported sample.

No matching public Catchpoint/WebPageTest result or exact-flag PSI lab report was retrieved. Public WebPageTest history searches over 30 days and one year returned no matching rows. This does not exclude private, unlisted, inaccessible or expired tests. No hosted or paid test was submitted.

PSI field data would reuse CrUX rather than independently replicate it. No URL-specific HTTP Archive numeric record was retrieved. PSI methodology · Catchpoint public-share documentation.

These are independent, phase-clipped diagnostics—not timing samples. The covered mobile pre-LCP interval is 75.4% sampled idle; desktop’s shorter startup interval is only 18.2% idle. The first 208 ms desktop / 422 ms mobile after navigation are not covered. Non-idle includes browser engine work, not just JavaScript.

Room samples are clipped from the trusted click through one second after the diagnostic frame, removing pre-click profiler overhead. Full profiles and URL-level buckets remain available, but their whole-window idle shares are not startup occupancy. Clipping method, intervals and source buckets.

URL-level leaf time is not a React commit duration, package owner or critical-path delay. Browser engine categories and unattributed work are not reassigned to React. End-only landing counters are not presented as measured CPU deltas.

Obtain matching build graphs and source maps, then correlate narrow navigation-to-LCP and click-to-frame traces with component commits. The public source-map lookup returned 404; private maps and a profiling build were not available. Public script-marker checks are limited substring evidence, not a dependency inventory.

React Server Components: a narrow delivery experiment

RSC can keep selected component work and code out of the client, but interactive client boundaries remain. Test a read-heavy slice, measuring client bytes/execution, HTML or Flight transfer, streamed readiness, server CPU/memory, TTFB and caching. Server-rendered HTML or createRoot alone does not establish deployed RSC. React’s RSC documentation.

Qwik: compare resumability with a fair baseline

Resumability changes startup work, but a rewrite also changes routing, chunking, prefetch, serialization and server execution. Compare it with an optimized React baseline using equivalent content, legal flow, player behavior, focus, state and cache conditions. No Qwik arm was measured here. Qwik’s resumability model.

Partytown: only for an eligible measured owner

Delete or defer unnecessary third-party work first. A worker can change DOM/event behavior, cookie access, CSP/CORS needs and request semantics. Local worker-message requests are not external HTTP transfers. Do not move legal logic, security challenges, app bootstrap, navigation, media or account flows merely because a sampled URL looks costly. Validate one noncritical script, event parity and rollback. Partytown’s documented trade-offs.

Browser TTFB mixes client, network, connection, edge, cache and origin time. It cannot establish origin CPU or cost. RSC, SSR and server-side loaders may move rather than remove work.

A fair architecture test needs source-aligned server spans, CPU/memory, response bytes, cache-key correctness, cold/warm paths and errors, in an authorized test environment. None of those private origin measurements or a load test was obtained. Reject an apparent client win if it causes unacceptable server, accessibility or reliability regressions.

The analysis includes twenty valid observations, four separate CPU diagnostics and three sensitivity checks. One 58.6-second desktop landing recording is excluded for its long window; three room recordings are excluded because their orphan network events lacked phase classification. Their timing observations are not described as playback failures.

Landing observation windows span 30.315–30.607 seconds; room windows are approximately 30.55–30.63 seconds. Observers were installed before every recorded FCP and report no drops or resource-buffer overflows. Installation still follows a new-document poll, not document-start injection; absence of recorded drops is not a completeness guarantee. Performance Timeline buffering rules.

Network coverage is parent-target only. Encoded completion bytes omit unfinished partial transfer. The legacy CDP network-shaping command is recorded and deprecated; CPU slowdown is host-relative and uncalibrated. Live content, one host, retained storage, browser extensions and observer overhead remain part of the environment.

No performer images, media payloads, descriptions, chat, account data, headers, cookies, response bodies, incidental request query strings or function names are saved. The target URL, public script basenames and sanitized geometry are retained for attribution.

Experiment assignment, private build graphs/maps, source-aligned component traces, physical-device replication, origin CPU and matched outside synthetic results remain unresolved. These are explicit limits, not completed tests.

Measurements and per-run evidence · Supporting-file integrity manifest.

• Chrome Network events: request starts, outcomes and completed encoded bytes; parent target only.
• PerformanceObserver: paint, layout shifts and tasks; bounded observer coverage.
• Video-frame callbacks: source-qualified frame observations; not physical display.
• CPU profiler: sampled execution, clipped to stated intervals; not package attribution.
• CrUX: field context; query variants and same-document journeys aggregated.
• Integrity hashes: file consistency; not experimental causality.

Units: MB means 1,000,000 bytes; MiB means 1,048,576 bytes.

The offline summary script filters HTTP(S), preserves missing values and computes nearest-rank p50/p75. Every five-run p75 is simply the fourth ordered observation, not a field percentile. Hashes establish file consistency, not causal correctness.