chore(examples): add bundle analysis and blocking size budgets - #1145

Why

The bundled example packages had no way to tell what ends up in their bundle, and no guard against it growing.

The Vite examples declared a chunkSizeWarningLimit, but it only makes Vite log a warning and the build still exits 0, so a bundle regression showed up as a single line in a CI log that nobody reads (upstream request: vitejs/vite#18496). The webpack examples had no budget at all.

What

Bundle analysis, opt-in, in the six bundled examples. Rsdoctor for the three webpack examples, vite-bundle-analyzer for the three Vite ones, since Rsdoctor supports only webpack and Rspack. Both are enabled by a dedicated npm run build:analyze script, which sets ANALYZE=true through cross-env so the command also works in Windows shells. The regular npm run build is untouched: no analyzer, no report server, no source map. A source map is generated for the analyze run only, hidden-source-map for webpack and sourcemap: 'hidden' for Vite, so the analyzed bundle stays byte for byte the one the regular build produces, content hash included; Rsdoctor needs it to attribute bytes to individual modules.

The production build now fails when a bundle grows. The webpack examples get a performance budget with hints: 'error', in production only, since development bundles are not minified and would break npm run dev. Vite cannot do this on its own, so the examples pass their limit to a shared plugin, scripts/vite/chunk-size-limit.mjs, which fails the build naming the chunk and both sizes. Each limit is declared once per example and feeds both chunkSizeWarningLimit and the plugin, so the warning and the error cannot drift apart.

Each value is the smallest one that passes, that is the current size rounded up to the next kB. The intent is to follow the size evolution rather than to leave room to grow into, so the incidental headroom is accepted; the convention is recorded in .claude/rules/tooling/bundle-size-budgets.md.

--fail-at-end in scripts/build-all-examples.bash, used by the CI. Once a build can fail on size, the first failure would abort the script and hide both the other examples and the size table, so getting the full picture would take one CI run per failing example. With the option, every example is built, the file listing, the markdown table and the CSV are still printed, and the failures are reported after them. The default behavior is unchanged: without the option the script still stops at the first failing build.

Shared Vite configuration. The codeSplitting group that isolates @maxgraph/core in a dedicated maxgraph chunk was copy pasted in the three Vite examples; it now lives in scripts/vite/maxgraph-chunk.mjs next to the size check it feeds. The three vite.config.js are identical apart from their limit, and the three webpack.config.js apart from their budget.

Dependabot. New vite group for vite and vite-*, and @rsdoctor/* added to the existing webpack group, so those libraries move together across the examples. @storybook/html-vite is deliberately left in the storybook group, which owns the Vite version Storybook builds against.

Sizes and budgets

No bundle size changed in this PR. The js-example-without-defaults budget was initially calibrated against a stale packages/core/lib; rebuilding core brings that bundle to 239,112 B, and the budget follows.

Decisions worth flagging

• The analyzer and the size guard are never registered together in the Vite examples. vite-bundle-analyzer is a post plugin, so the guard's error aborted the build before any report was written, exactly the case where the report is needed to find out what grew. npm run build, which the CI runs, remains the command that enforces the limit.
• --list-size-only --fail-at-end is a documented no-op, not an error: nothing is built, so nothing can fail, and a caller passing flags generically should not be rejected for a harmless combination.
• The artifact upload now runs with if: ${{ !cancelled() }}. Skipping it on failure would defeat the purpose of --fail-at-end, since the dist of the examples that did build is what one wants to inspect. This does not weaken the gate: the job still fails, so the jobs depending on it, the release included, do not run.
• A failed Vite build leaves no dist, so a failing example disappears from the size table and the CSV instead of showing a stale number.
• Budgets are tight by design, between 36 B and about 900 B of headroom. A bundler patch bump can therefore turn a build red; the fix is to check the increase is intended and update the value in the same commit.

Validation

• ./scripts/build-all-examples.bash and ./scripts/build-all-examples.bash --fail-at-end: exit 0, the six examples below their limits, no webpack performance warning and no Vite chunk warning.
• Guard proven blocking: with a limit lowered to 10 kB, RolldownError: Chunk "assets/maxgraph-Be5ReXNy.js" is 361.55 kB, above the 10 kB limit. and exit 1.
• --fail-at-end proven: with two limits lowered, both failures are listed after the size table, the third example still builds, exit code 1.
• Analyze runs verified on both bundlers, report reachable over HTTP, and the emitted bundles unchanged to the byte with identical content hashes.
• The shared Vite module resolves from the three example packages, and the extraction is a pure refactoring: same chunk sizes and same content hashes as before.
• npm run lint clean. Note it only covers **/*.ts, so the two new .mjs files are outside its scope; they were linted explicitly.

Summary by CodeRabbit

• New Features

  • Added local production bundle analysis for Webpack and Vite examples.
  • Added production bundle-size checks that fail builds when configured limits are exceeded.
  • Improved dedicated maxgraph bundle generation.
  • Build scripts now process remaining examples, report failures, and summarize results.

• Documentation

  • Added guidance for bundle analysis, size budgets, troubleshooting, and updating limits.

• Chores

  • Improved dependency grouping and build artifact handling in automation.