[tbvjaos510]

This is the Vite side of a two-part fix; the rolldown half landed in rolldown/rolldown#10386 and shipped in rolldown 1.2.1, which main already pins.

What this solves

In bundled dev mode, a lazily compiled chunk is served from /@vite/lazy?... and ends with a //# sourceMappingURL=lazy_compile_N.js.map reference — but that .map was never registered anywhere, so the browser gets a 404 and the module can't be mapped back to its source.

Eager chunks (storeOutputFiles) and HMR patches already register their maps in memoryFiles; lazy chunks were the only path missing it. This registers the lazy chunk's map the same way.

The reference is relative and the chunk is imported from /@vite/lazy?..., so it resolves under /@vite/. The map is registered under that directory for the reference to reach it — no rewrite of the emitted chunk is needed.

rolldown#10386 is the companion fix: it makes the map's content correct (it previously mapped to post-transform positions instead of the source the user wrote) and makes compileEntry actually return the map (sourcemap / sourcemapFilename). With both, lazy chunks get correct, reachable sourcemaps under Vite's default output.sourcemap = true.

Alternatives explored

• Inline the map as a data: URL (Vite's classic dev approach). Rejected: eager chunks and HMR patches in bundled dev are served as separate .map files, so inlining would make lazy chunks the odd one out. Keeping the separate-file approach is consistent with the rest of bundled dev.
• Rewrite the chunk's sourceMappingURL to an absolute /lazy_compile_N.js.map. Works, but depends on a fragile string replace of the exact comment format. Registering the map under the /@vite/ directory instead needs no rewrite and keeps rolldown's emitted reference untouched.

Reviewer attention

• The @vite/ prefix on the map key mirrors the /@vite/lazy?... import path that rolldown's proxy template hardcodes and that the lazy-bundling middleware already matches on. It isn't extracted to a shared constant — that string is currently duplicated (middleware + here) with its source of truth in rolldown; happy to pull it into a constant if preferred.
• The test lives in the existing hmr-full-bundle-mode playground: the lazy bundling test now also asserts the map is served (200) and maps back to dynamic.js. It fails without this change (the reference 404s, and the map fetch returns the HTML fallback).

🤖 Generated with Claude Code

[btea]

@tbvjaos510 It seems that there is a conflict in the code that needs to be resolved.

[tbvjaos510]

@btea Thanks! Rebased onto main

[h-a-n-a]

@btea I don't have the permission to merge. Hand it over to you. Thanks 🩵

[btea]

I don't have it either. Let's leave it to others. 😄