Description

The hmr-whole-chain-dispose browser fixture runs on Vite's full-bundle client (the provisioned vite/ checkout, vitejs/vite rolldown-canary), not rolldown's native runtime — see the note in full-bundle-dev-environment.ts. That client's handleModuleCacheRemoval was recently fixed to preserve import.meta.hot.data across a module's HMR generations, matching Vite's documented contract:

The import.meta.hot.data object is persisted across different instances of the same updated module.
— https://vite.dev/guide/api-hmr#hot-data (see vitejs/vite#23001)

On rolldown-canary the client (renamed to bundledDevHmrClient.ts) now does:

handleModuleCacheRemoval(id: string): void {
  const disposer = this.disposeMap.get(id)
  if (disposer) {
    disposer(this.dataMap.get(id))   // existing bag, not a fresh {}
  }
}

But this test + fixtures still asserted the old "each generation gets a fresh data bag" behavior, so once the client started preserving, CI failed:

FAIL playground/hmr-whole-chain-dispose/__tests__/hmr-whole-chain-dispose.spec.ts
  > every re-executed module disposes, each generation gets a fresh data bag
AssertionError: expected 'baz-own-write' to be null

('baz-own-write' is a direct hot.data write from generation 0 that correctly survived into generation 1.)

Change

Align the fixtures and assertions with the preserved-data contract:

• Rename the leak probe (which meant "a direct write that must NOT survive") to own, and assert each module's own direct hot.data write is read back by the next generation (own === \${id}-own-write``).
• gen0 bags are still empty (gen: null, own: null); the disposer-written gen counter still advances 1 → 2 — those were already passing and are unchanged.
• Whole-chain dispose coverage is untouched: every re-executed module in the evicted chain (baz, bar, and the boundary foo) still disposes each generation (3, then 6).

Validation

• vp check (lint/format) via the commit hook.
• The browser E2E itself is verified by CI (node-dev-server-test); it requires the provisioned vite/ checkout + native build, so it was not run locally. Expectations were derived from the preserved-data semantics and confirmed against rolldown-canary's current bundledDevHmrClient.ts and the exact CI failure value.

Note: the separate Vite Test job failing on the same runs is the Vite-side manifestation of the same behavior change and is not addressed here (it lives in the vite/ checkout, not this repo).