Testing and experiments

Run the checks

dotnet restore --locked-mode
dotnet test -c Release --no-restore
dotnet publish src/WebClient -c Release --no-restore -o artifacts/publish
dotnet publish src/WebClient.Wasm -c Release --no-restore -o artifacts/wasm
dotnet list package --vulnerable --include-transitive

Tests use xUnit v3 (the xunit.v3.mtp-off package, so dotnet test keeps running them through VSTest), bUnit, ASP.NET Core’s in-process server, and real SQLite migrations. Test databases live in task-owned directories under the test build output and are removed after the test host shuts down.

What the tests prove

Layer	Behavior
Components	Draft cancellation/confirmation, required-field rejection, callback/reset behavior, preference write paths, storage failure fallback, unsaved self-closing drawer
Forecast services	Deterministic IDs and paging, validated commits, CSV quoting, bounds, lossless escaping and blank lines, culture-aware search, cancellation
HTTP integration	Prerendered headings on every lab, liveness/readiness, bounded API requests with validation problems, cancellation and simulated failure, security headers, path-preserving skip link, base-relative links
Identity	Antiforgery rejection, anonymous/student/instructor policy results, disabled demo sign-in
Persistence	Migrations, workspace isolation, stale-update/delete conflicts, scoped reset, note limit, page-only sliding workspace cookie, tampered-cookie replacement
Static demo	In-browser API rules, encoded typed-client requests, localStorage notebook outcomes, browser personas, existing lab sources
Gallery	Every component page and page sample renders with bUnit without hitting an example’s error boundary, every example embeds its source, routes are unique, the highlighter round-trips source, snippets omit defaults
Localization	Culture cookie, Portuguese resource text, culture-specific formatting

These checks do not measure browser layout, assistive-technology behavior, or a live hosting environment. The WebAssembly demo has a browser check (below); the server app’s browser behavior and post-deployment verification remain separate layers. Hostinger setup is pending.

Check a published app through HTTP

Run the published app in one terminal:

dotnet artifacts/publish/WebClient.dll \
  --contentRoot "$PWD/artifacts/publish" --urls http://127.0.0.1:5000

Then:

python3 scripts/smoke-http.py --base-url http://127.0.0.1:5000

The script retries readiness, checks that every route prerenders its own heading, retrieves required static assets, and reads one API page. It does not click the UI.

Check the WebAssembly demo in a browser

Build the Pages site with the demo, then drive it in Chromium:

dotnet publish src/WebClient.Wasm -c Release -o artifacts/wasm
dotnet run --project src/WebClient.Prerender -c Release -- artifacts/wasm/wwwroot
cd docs
npm run build
python3 ../scripts/prepare-pages-demo.py
npx playwright install chromium   # once; or set PLAYWRIGHT_CHANNEL=chrome
npm run check:demo

The check serves docs/out like GitHub Pages (repository path, .html for extensionless paths, 404.html fallback) and verifies that deep links arrive prerendered with their own title, that the app takes over, that the demo starts offline after one visit, quick search, playground links, the in-browser API, the skip link, scoped state, notebook conflicts, browser personas, culture reload, atomic CSV import, the largest grid, the not-found page, and a clean console. It then visits every routable page of the shared library inside the running demo — every lab, component page and page sample — and fails if a page lacks a single heading or an example hits its error boundary.

Accessibility audit

On every demo page, and on the landing and guide pages, the check runs axe-core with reduced motion and fails on any violation of the rules our markup controls: contrast, links distinguishable without color, unique landmarks, labels and accessible names, names that contain their visible text, and focusable elements hidden from assistive technology.

MudBlazor 9.11 renders a few elements that a page cannot label from outside. They are excluded by selector so the audit stays meaningful. Each is reported upstream:

Excluded	Why	Issue
`.mud-treeview .mud-checkbox-input`	Multi-selection tree-view checkboxes have no accessible name	#13977
`.mud-picker-color-content`	The color picker’s channel inputs have no names; swatches put `aria-label` on role-less elements	#13976
`.mud-table-pagination`	The rows-per-page select is not named by its caption	#13978
`.mud-table` menu activators and loading bar, grid header icon buttons	The data grid’s settings menu, column drag handles and loading bar are unnamed	#13979
`.mud-nav-group nav:not([aria-label])`	MudNavGroup wraps its items in a second, unnamed `<nav>`	#13980
`.mud-nested-list`	Nested lists render a listbox inside a listbox and `aria-expanded` on options	#13981
`.mud-avatar-text`	`Variant.Text` ignores `Color` and renders white on light grey	#13982
`.mud-popover-provider`	Popover content renders outside the page landmarks	—
Duplicate `aria-label="Breadcrumb"`	MudBreadcrumbs names every `<nav>` “Breadcrumb”; a passed `aria-label` lands on the inner `<ol>`	#13974

Sliders and radios are not excluded: MudSlider does not link its visible label to the range input (#13973), so every example passes an aria-label that starts with the visible text. A radio’s custom content must not use MudStack, whose role="group" hides the text from the radio’s name.

Other library issues the audit does not cover: positive tabindex in the carousel (#13975), a focusable wrapper around a custom menu activator (#13983), and chips inside a select nested in an interactive element (#12613). The pages document the workarounds.

Compare performance honestly

Keep the same machine, build configuration, seed, and dataset size.
Warm up once and record several runs.
Separate generation time, server filtering time, network payload, and browser rendering.
Compare the in-memory grid with the paged API using the same dataset size.
Try ReadyToRun with docker build --build-arg READY_TO_RUN=true ... and compare startup plus image size against the default image.

The grid’s displayed timing covers generation only. Virtualization is not a server memory limit. A stress dataset demonstrates a tradeoff; it is not a concurrency benchmark.

Suggested manual UI pass

Navigate using only the keyboard, including copy and dialog actions.
Edit, cancel, reopen; confirm the original value remains.
Submit empty and invalid fields; inspect focus and validation feedback.
Switch themes at desktop/mobile widths and inspect contrast.
Switch culture, reload, and verify selected language and formatting.
Cancel API and file work; navigate away during a delayed request.
Open the notebook in two tabs and provoke a conflict.
Disconnect/reconnect and observe circuit ownership.

Add the smallest meaningful test

Test the observable contract at the lowest sufficient layer. Keep one parameterized test when it can prove several related outcomes. Add a distinct case for an independently regressing boundary, such as a stale database version or a rejected oversized import.