Companion repository to rolldown/rolldown#9441.

Best-effort scaled reproduction — does not cycle on its own at this scale. This repo mirrors the failing app's structural patterns (1,121 source files, 120 React.lazy routes, mixed CJS/ESM vendor deps, legacy + modern code, tss-react withStyles with lodash/<sub> defaults, import.meta.glob lazy routing, multi-entry HTML with overlapping lazy sets) but rolldown's default chunker successfully avoids cycles for this graph.

In the real-world failing app (~5,000 modules, similar deps and patterns), vite 8.0.13 (rolldown 1.0.1) produces 8,678 chunk cycles including the production-failure shape — and vite 8.0.11 (rolldown 1.0.0-rc.18) on the same source produces 0 cycles. So the same graph crosses the chunker's threshold at production scale that this synthetic does not.

• Production-realistic Vite 8 + React 19 setup with @mui/material, @mui/x-data-grid, @mui/x-date-pickers, @tanstack/react-query, @emotion/styled, tss-react, react-hook-form, react-router, zustand, yup, lodash (main + subpath), clsx, date-fns

• 30 entity modules × {components, hooks, utils} + 60 legacy components in cyclic legacy containers + 50 shared "glue" components + 40 shared MUI wrappers

• 120 lazy routes via import.meta.glob and two HTML entries (index.html, page2.html) with overlapping lazy sets

• A barrel.tsx that re-exports many shared utilities and references legacy components — produces a barrel-<hash>.js shared chunk in the output graph

• A jsx-runtime-<hash>.js extracted chunk (matches the shape from rolldown's own #9331 regression)

• Every Route chunk has the exact emission pattern that fails in production:

  import { S as e, b as t, t as n } from "./jsx-runtime-<hash>.js";
  import { ... } from "./barrel-<hash>.js";
  var g = e(t(), 1);           // ← this shape fails when t is in a TDZ cycle

  In the failing app t reads back as undefined because the chunk graph contains a cycle that puts it in a not-yet-initialised state at module init. In this synthetic the chunk graph is acyclic, so t is initialised by the time the call runs.

npm install
npm run build         # vite 8.0.13 + rolldown 1.0.1
npm run preview       # http://localhost:4173

Smallest cycle observed:

big-vendor.chunk.js  (1.7 MB, 1196 sources — lodash + @mui + @tanstack + ~700 app sources)
  imports "t as Wd" from
mid-legacy.chunk.js  (34 KB, 4 sources — two legacy form components + their styles)
  imports "t as w" from
button.chunk.js      (9 KB, 2 sources — legacy <Button1> + its styles, uses lodash/get)
  imports "T as c" from
big-vendor.chunk.js  ← cycle closes back

T as c is lodash/get's __commonJSMin wrapper. The cycle puts that var binding in TDZ-equivalent state when the button chunk's top-level var d = a(c()) runs, so c reads undefined → TypeError: c is not a function.

• The #9331 fixture verbatim
• 3-way util cycle imported by N lazy routes
• Source-level .cjs cycles with top-level cross-chunk calls
• 100-module shared registry pulling many lodash/<sub> paths
• manualChunks forcing the cycle (works — but builds the same on rc.18, so it's not the regression)
• Multi-entry vite with overlapping lazy sets (this repo's current state)
• import.meta.glob lazy routing (this repo's current state)
• barrel.tsx JSX component to force a barrel → jsx-runtime edge

The chunker is robust at every shape I could construct at ~1,000-file scale.

The chunker change in vite 8.0.12 → 8.0.13 (rolldown 1.0.0 → 1.0.1) came from two PRs:

• v1.0.0 #9270 — implement dynamic dominator merge logic
• v1.0.1 #9305 — chunk-optimization: dedupe already-loaded dynamic deps

#9305 covered the #9331 case (confirmed — fixture builds cleanly). The hypothesis is it didn't cover an adjacent case the production graph hits, where the "intersection of loaded atoms from every importer entry" check leaves a var-bound __commonJSMin wrapper TDZ-exposed across the cycle.