A Turbopack production build that reused .next after the mobile merge
emitted CSS without the new touch-target rules of globals.css, so NAV-038
measured 36 px header buttons that the source makes 44 px. A clean build
has the rules. Acceptance runs now remove .next before building, so a
gate always tests the source as it is.
Phone and tablet coverage (#708) on the v1-only line:
- mobile's phone-layout test changes keep the integrated v1 routes; the
mobile-added phone tests that waited on unversioned routes now wait on
the v1 ones (query-phone bulk edit init, adornment download links)
- the TABLE_BASED "unsupported authentication" test (SEC-032) stays
removed: TABLE_BASED sign-in is supported on this line (#700)
- saved-report tests keep the create-child dialog flow with mobile's
openRecord/recordAction helpers and @mobile tags
- processRecords accepts an omitted records list only for a run with no
records (mobile's rule) and keeps the v1 empty-list handling
- DataGrid tests keep both the row-rendering and the checkbox-label suites
- matrix rows renumbered clear of the integrated line: NAV-030->NAV-037,
NAV-031->NAV-038, REC-049->REC-058, REC-050->REC-059 (matrix, test
titles, browser-matrix.md)
- docs/acceptance/browser-matrix.md takes the phone and tablet scope
draft, with the renumbered rows and the integrated TABLE_BASED, login
branding and record search rows
RapiDoc, which the table developer view loads on demand for the API docs,
is a 204.5 KB gzip chunk. Counting it in the application's total and
largest-chunk budgets would hide every other regression behind one raise.
perf-budget.json now lists it under lazyChunks: the script finds the chunk
by its custom-element definition, fails if any route loads it on first
load, holds it to its own 215 KB budget (about 5% above the measured
size), and leaves it out of the total (702.3 of 710 KB) and largest-chunk
(93.8 of 100 KB) budgets, which are unchanged.
- scripts/perf-budget.mjs (pnpm perf:budget): measures the static
export (out/) per route: every script and stylesheet the route's HTML
loads before hydration, gzip -9, plus total JS and the largest chunk
(Next 16 no longer prints these). Compares them with perf-budget.json,
writes test-results/performance/bundle.json and exits 1 on a breach.
- perf-budget.json: budgets about 5% above the measured sizes.
- CI: a "Bundle budget (static export)" job runs build:export and the
check on every PR.
- Worst regression found: /app/[slug] (app home, record query, process,
report) imported AppHome from the widgets barrel, which re-exports the
chart widgets, so Recharts + d3 (a 94 KB gzip chunk plus 7 KB) was in
the first load of every table and process page although WidgetRenderer
loads charts lazily. The slug and saved-view pages import their page
components by module path now. /app/[slug] first-load JS:
562.9 -> 448.2 KB gzip (-20%); total JS 689.2 -> 675.0 KB, 54 -> 51
chunks. Charts still load on demand on dashboards.
Adds an iPad (WebKit, touch) acceptance project that runs the @mobile and
@tablet tests, sizes controls to 44 x 44 CSS px on coarse pointers, and makes
the gate require a passing touch-project test for every required row unless
the row records why it is desktop-only.
- .github/workflows/ci.yml: three parallel jobs on pull_request and
pushes to main/develop, none needing a backend:
unit (typecheck, lint, test:coverage), e2e (pnpm build, then the
mocked Playwright suite against the standalone server, chromium) and
storybook (build-storybook).
- Mocked e2e reliability:
- saved-views.spec mocked /qqq/v1/processes/<name>/init with a JSON
`values` field, but the app posts saved-view processes to the
registered route (/processes/<name>/init) with one multipart field
per value, so three tests saw no views. The mocks now match the
real request.
- playwright.config: QQQ_E2E_SERVER=production serves the `pnpm build`
standalone output (scripts/e2e-server.mjs) instead of compiling
routes on first hit; QQQ_E2E_PORT picks the port; retries 0 so a
flaky test fails instead of passing on retry; trace kept on failure.
- 104/104 pass; 520/520 with --repeat-each=5 --workers=4.
- vitest.config: coverage thresholds for src/lib just under the measured
82.8/75.65/89.87/84.42 (statements/branches/functions/lines), and
higher per-folder floors for src/lib/api and src/lib/utils.
- Storybook: silence Rollup's per-module "use client" directive and
sourcemap-location warnings (100 lines per build).
- playwright-report/index.html and test-results/.last-run.json were
committed despite being ignored; untracked.
- The mobile (Pixel 7) project runs the tests tagged @mobile: navigation
drawer (NAV-026, INT-008), card list (INT-008, INT-009), record view,
forms (INT-007, PRC-009), process runs (PRC-001, PRC-004), dialogs
(INT-003), sign-in/out and expiry (SEC-020..022) and the axe scan
(INT-005). Desktop-only layouts are not phone scenarios.
- New INT-009 phone test: a held list shows busy placeholder cards, not
'No records found' (QRun-IO/qqq#694).
- INT-005 and SEC-021 accept the card list where they checked the grid.
- The gate reports per-project counts and fails a project that ran no
tests.
Static export mode with placeholder routes and Maven packaging for the Javalin-served default dashboard, plus the matrix-gated acceptance suite (QRun-IO/qqq#649).
- Add scripts/check-license-headers.mjs — walks src/**/*.{ts,tsx},
detects missing headers, exits 1 in CI; --fix mode auto-prepends
- Add .githooks/pre-commit — runs pnpm lint:license before every commit
- Register .githooks via git config core.hooksPath in package.json prepare
(runs automatically on pnpm install, zero extra dependencies)
- Update pnpm lint to include license check — enforced in every CI pipeline
- Add pnpm lint:license and pnpm lint:license:fix convenience scripts
- Apply Apache 2.0 header to all 227 existing src/ files
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>