[santhiprakash]

What is this PR solving?

Fixes #23296

When using the oxc minifier (Vite 8 default, introduced by #22468), legacy chunks are minified with module: true because rolldown derives this flag from output.format (which remains esm at minify time). The legacy plugin then converts chunks from ESM to SystemJS in renderChunk, which runs after minification.

SystemJS output runs in sloppy mode (no "use strict"), where Annex B.3.3 hoists block-level function declarations to the enclosing function scope. With module: true, the minifier applies strict-mode semantics and can dead-code-eliminate block-scoped functions that are actually reachable via sloppy-mode hoisting. After mangling, this causes name collisions that produce TypeErrors and white screens on legacy browsers.

What changes did you make?

• packages/plugin-legacy/src/index.ts: Disable rolldown's built-in minification for legacy output chunks (it always uses module: true). Instead, invoke minifySync from rolldown/experimental directly in renderChunk with module: false, compress: { target: 'es2015' }, and mangle: true. This ensures the minifier uses sloppy-mode semantics for SystemJS output. Resolves rolldown through vite's dependency tree since it's not a direct dependency of plugin-legacy.

• packages/plugin-legacy/src/__tests__/minify-systemjs.spec.ts: New unit tests verifying that minifySync with module: true eliminates block-scoped functions (strict-mode DCE) while module: false preserves them (sloppy-mode safe), and that compress: { target: 'es2015' } doesn't introduce newer syntax.

• playground/legacy/__tests__/legacy.spec.ts: Updated the minification heuristic from terser-specific patterns to SystemJS format detection, and added a polyfills-legacy assertion.

Alternatives explored

• "use strict" injection (closed-unmerged PR fix(legacy): emit legacy chunks as strict mode code #23297): Rejected by maintainer sapphi-red as incomplete — the directive alone doesn't fix the DCE/mangling behavior.
• Output-level minify: { module: false }: Not possible by type — rolldown's OutputOptions.MinifyOptions omits the module field; it's derived internally from output.format.
• minify: 'dce-only': Also uses ESM module semantics internally, so it has the same DCE problem. Must use false (no built-in minification).

Testing

• 3 new unit tests in minify-systemjs.spec.ts pass (direct minifySync behavior)
• All 30 legacy playground e2e tests pass (including updated minification heuristic)
• All 34 plugin-legacy unit tests pass

The new sloppy-hoisting playground test fixture from earlier iterations was removed because the block-scoped function pattern gets eliminated by rolldown's bundler-phase tree-shaking (not just minification), making it unsuitable for e2e. The unit tests cover the minify-level fix directly.

AI disclosure

This PR was developed with AI assistance. I reviewed all generated code, verified each change against the root cause analysis, ran all tests, and validated the fix approach against the maintainer's guidance (sapphi-red's feedback on #23297).

[santhiprakash]

Rebased onto current main (a98c8d95d) to clear the merge conflict — only conflict was in packages/plugin-legacy/src/index.ts import block (upstream mainline moved @babel/core/browserslist into the alphabetical stdlib block; kept my added pathToFileURL named import + functional changes intact). playground/legacy/__tests__/legacy.spec.ts auto-merged. New head: 7982b9a9. CI rerun in progress.

[santhiprakash]

Follow-up fix on top of 7982b9a9: the rebased head failed the CI Lint job (pnpm format && git diff --exit-code) because .oxfmtrc.json enables sortImports and the new minify-systemjs.spec.ts had the vitest import before rolldown/experimental. Swapped the two lines in packages/plugin-legacy/src/__tests__/minify-systemjs.spec.ts so rolldown/experimental comes first. Pure formatting, no logic changes. New head: 88bbf40f. Lint/Build&Test on Linux/macOS pass; the only remaining failure is a Windows Test serve flake in playground/environment-react-ssr > pre-bundling > deps reload (timing-sensitive test that fails on the Windows runner on mainline too — unrelated to the plugin-legacy changes).

[github-actions]

This PR has been automatically flagged as likely to be created by a bot, LLM, or agent, and will be automatically closed. These contributions harm the maintenance of the project. Please read our AI policy for more information.
If you believe this is a mistake, please reply to this comment and we will review it.

[santhiprakash]

Hi, the close/reopen above was a manual CI retrigger after the Windows Build&Test job flaked, because the santhiprakash account cannot rerun only the failed job on vitejs/vite. The branch and diff are unchanged. Could a maintainer please add the bot: skip label and reopen so the actual CI status can be evaluated? I can explain the legacy-minify diff and respond to review feedback.