# Browser performance audit — November 2, 2026

Canonical report: https://kitsuvo.com/browser-performance

## What was run

Apple M5 Max, 64 GiB RAM, arm64, Darwin 25.6.0. One shared host; other host activity not fully controlled. Hidden headless windows, 800 × 560 CSS pixels at DPR 2, fresh disposable profiles. No account sessions, model calls or real external test websites. Browser startup/profile creation is outside timed observations. Defaults differ: Chromium background networking is disabled by its launch flag; Kitsuvo's fixture profile uses Local only, automatic updates off, and filtering off to keep all fixture content accessible. This is not an out-of-box experience comparison.

Versions: unreleased Kitsuvo candidate 0.2.38 (application source 35d5a4ca3a7e0e203de63c0876fc8ae44bfc92e2), Chrome 156.0.8078.76, Edge 155.0.4283.45, Firefox 157.0.1. Binary hashes, browser versions and helper/harness hashes are in manifest.json. The fresh native production candidate passed its build/exposure/Plus checks; this does not establish public availability or native Windows/Linux/mobile performance.

Two sequential rounds in reversed order. Fifteen warm timed observations per browser and workflow per round. Three untimed warm-ups for evaluations/snapshots/form tasks, two for navigation. Both rounds pooled without trimming. Nearest-rank median and p95 from thirty observations; tail estimates remain sensitive to individual observations. All 1,680 primary round-trip observations passed their output assertions.

The identical 42,743-byte helper replaces the helper preinstalled in Kitsuvo. Full and bounded snapshot output hashes match across all four browsers. Viewport-only outputs differ across engines because native layout differs: for the table, Kitsuvo returned 20,307 characters, Chromium browsers 20,100, and Firefox 19,893. Viewport rows are not equal-output timing comparisons. Three fixtures: small form, 1,800-paragraph article, 1,500-row table. Full snapshots cap at 30,000 characters, bounded snapshots at 2,000; full means the configured snapshot variant, not an unbounded complete-document read. Navigation verifies the expected heading. Form edits verify the actual field value and subsequent snapshot.

Page-work timing isolates helper execution. Timer precision differs by engine; zero means below the timer bucket, not zero work. No sub-millisecond speedup ratios are claimed. Navigation and evaluation round trips include different native-control/CDP/WebDriver transports and readiness semantics. They are not pure engine or competitor MCP-product comparisons. Firefox NavigationTiming and raw per-round observations are retained alongside all other results. Published Firefox files correct one metadata label from native-load-state to webdriver-complete (the report ternary omitted the newly added adapter). A metadata_correction record retains the original value/reason. No samples or outcomes were changed. The committed harness also fixes that metadata ternary; manifest.harness_sha256 identifies the measured harness before this metadata-only correction.

## Network, weight, CPU and memory

Network: loopback server response-body bytes, including setup, warm-ups and favicon probes. The simple fixture server returns HTML for every path, including favicon probes. Extra favicon and adapter-priming requests explain different totals; these are not bandwidth or data-saving rankings. Headers, protocol traffic, background connections and external transfers are excluded.

Weight: allocated app-directory size measured with `du -sk`, in KiB. Public Kitsuvo 0.2.37 is a different source from the runtime candidate. Public archive SHA256: 2fe06929ad9483df5c39ef42a8de4c07c6c7307f2458ab496df1aae74ba6124e, source deb481e122897b4b65999e79a43efebdc442983b. Public binary SHA256: a939a26b64c0f0c74f0314c5875d20ffeea153fed466842f6b3fec52d30e71db. The minimal candidate app bundle is intentionally excluded from the size table. Profiles, models, downloads, system/shared frameworks and filesystem sharing effects are not included. macOS supplies Kitsuvo's WebKit; do not describe app-directory size as total footprint. Safari's system-managed bundle was not comparable under this measurement.

Resources: `ps` RSS at beginning/end, cumulative main-process CPU-time delta and elapsed wall time over the workload. Renderer, WebKit XPC, GPU and network helpers are excluded. These numbers cannot establish total-browser memory/CPU efficiency. No energy, battery, scrolling frame-time, cold-start or real-network ranking is established.

A distinct candidate-only native soak ran 12 tabs, two analysis bursts and 60 seconds of repeated navigation/form/snapshot work (27 cycles); all five correctness/cleanup checks passed. Its main-process RSS/idle CPU observations are in candidate-soak.json. It is not a leak test, sustained-stability claim, or cross-browser comparison. Fixture profiles were removed after successful runs; failed admission profiles were retained locally.

## Excluded results

Public Kitsuvo 0.2.37: extracted app control endpoint did not become ready within 30 seconds. No runtime score. Installed local 0.2.37 had a different source and was not substituted for the public build.

Brave 155.1.97.56: initial matrix reached its 180-second harness deadline. Bounded admission retry reached the debugging endpoint but failed with `CDP WebSocket open timed out`. No score. Safari 26.6.2: existing adapter is headed only, excluded from the hidden/headless comparison.

Failures are retained in admission-failures.json. They are harness/admission failures, not proof these products cannot browse normally.

## Reproduction in the Kitsuvo source checkout

No browsers or drivers are downloaded by these commands. Supply your staged, production-built Kitsuvo candidate and an existing geckodriver. Adjust installed browser paths in the runner if needed. The runner uses macOS app directories and is not a cross-platform native qualification command.

```sh
KITSUVO_AUDIT_BINARY='/path/to/Kitsuvo Candidate.app/Contents/MacOS/kitsuvo' KITSUVO_GECKODRIVER='/path/to/geckodriver' node scripts/run-browser-comparison.mjs target/your-audit
node scripts/publish-browser-audit.mjs target/your-audit target/your-published-audit
node scripts/benchmark-browser-workflows.mjs --browser kitsuvo --binary '/path/to/Kitsuvo Candidate.app/Contents/MacOS/kitsuvo' --mode headless --samples 5 --native-api on --metrics on --soak-seconds 60 --tabs 12 --burst-rounds 2 --output target/your-soak.json
```

The publisher is scoped to the November 2 four-browser protocol. A new environment/version/date requires a separately named record and metadata review; do not overwrite this historical dataset or label new measurements as this source. The CSV includes all aggregated timing rows; the JSON files retain individual observations. The local control token is held only in request headers and never published.
