Quality pass: library packaging, Job Monitor performance, state correctness, test coverage #494

Context

While working on the data-table performance refactor (#483), a full review of
diracx-web-components and the app shells surfaced problems in four areas. The
resulting work is too large to review as a single PR (+5000 lines), so it is
being split into a stack of focused PRs, all tracked by this issue.

What this covers

1. npm packaging of the published library

@dirac-grid/diracx-web-components currently only works inside this monorepo:

• No peerDependencies: react, react-dom, @mui/*, @emotion/*,
  @axa-fr/react-oidc are hard dependencies — consumers risk duplicate
  React/emotion instances ("Invalid hook call", broken styles) and a duplicated
  oidc module whose auth state silently breaks.
• next/image is imported in two components while Next.js is not declared as a
  dependency at all — the library is unusable outside Next.js apps.
• package.json#exports has no "." entry, so the root import advertised by
  main/module fails; no sideEffects: false, so tree-shaking and the lazy
  @mui/x-charts split are ineffective.

2. Job Monitor performance

• Every page/sort/filter change unmounted the table into a skeleton
  (no keepPreviousData); it now keeps stale rows visible with a progress bar.
• overscan: 50 and no memoized row component meant ~125 MUI rows
  (~1400 cells) re-rendered on every scroll frame and every state change.
• One monolithic context value: a single row-checkbox click re-rendered the
  search bar, the pie chart, and every status cell.
• Filter value suggestions fired uncached, unabortable jobs/summary group-by
  queries — several per restored filter, before any user interaction.
• The jobs search fetched all 18 columns while 8 are hidden by default
  (no parameters projection).
• Changing rows-per-page could serve minutes-old cached data
  (revalidateIfStale: false combined with a 60 s deduping interval).

3. State correctness and UX

• React state mutated in place in the dashboard import/creation flows and the
  search bar — latent StrictMode/concurrent-rendering bugs; one dialog opened
  only as a side effect of an unrelated re-render.
• Index keys on drag-reorderable lists (dashboard groups, filter equations).
• API responses and sessionStorage content trusted via blind casts — e.g. an
  empty job-history response crashed instead of showing the error snackbar.
• Token classification in the search bar raced the debounced suggestions
  effect: typing ID, =, 1 quickly silently dropped tokens.
• Assorted: sessionStorage writes during render in ThemeProvider, hardcoded
  light-mode focus colors in dark mode, inconsistent error surfaces, missing
  keyboard access and aria labels on the pie-chart legend and column toggles.

4. Test coverage and tooling

• Unit tests existed only for components with Storybook stories; the search-bar
  reducer, filter sync, and persistence hooks had none. The suite grew from
  60 tests / 7 suites to 93 tests / 11 suites.
• Two suites had vacuous or race-prone assertions; structuredClone was
  polyfilled with a lossy JSON round-trip.
• jsx-a11y linting was missing in the components package (where all the
  interactive UI lives); the app ESLint configs re-enabled formatting rules
  that conflict with Prettier.
• Hygiene: .nvmrc + engines, headless Cypress script, coverage script.