revert(runner): keep the compiler chunk's content hash (DEV-2569)

[demtario]

Follow-up to #249, from the high-effort review of it. Reverts only the hash-free compiler chunk name. The shape fix — asBabel at both import sites, which is what DEV-2569's second pass was actually about — is untouched.

The chunk is not self-contained

Rollup hoists the shared CJS interop helpers into the entry, so the emitted compiler chunk opens with a static import of a content-hashed file. Verified against the live deploy of #249:

$ curl -s https://demos.handsontable.com/assets/compiler-babel.js | head -c 60
import{c as SD,g as Nke}from"./index-BpieVTaX.js";

$ curl -s https://demos.handsontable.com/ | grep -o 'src="/assets/[^"]*"'
src="/assets/index-BpieVTaX.js"

and that entry chunk's top level is createRoot(document.getElementById("root")).render(…) plus Sentry.init (main.tsx: "Must stay first: initialises error reporting before any other module runs").

So in the one scenario the rename existed to cure — a tab open across a deploy that then triggers its first Tier-1 compile — the stable path resolves against the new build, statically pulls a 1.3 MB entry chunk the tab has never loaded, and evaluates a second complete copy of the app. React 18 clears the root container on mount, so the visitor's workspace is detached and silently remounted from a different build: unsaved edits gone, no card, two Sentry clients, a second copy of every module singleton including a second babelLoader latch.

That is strictly worse than what it replaced. The carded failure it removed says "Restart the preview to try again, or reload the page — a tab left open across a deployment has to reload to pick up the current version" — actionable, and it does not throw away work. Not losing unsaved edits is the stated reason rearmCompilerLoad exists at all.

The mirror case has the same root cause: a cached stable compiler-babel.js against a newer page carries a static import of a rotated-out ./index-<oldhash>.js, so the SPA fallback answers 200 text/html and the original DEV-2569 failure is back — now reachable on a fresh load rather than only in a stale tab.

The hash is load-bearing: it is what makes a rotated chunk fail loudly instead of resolving against the wrong build. #249's new e2e could not have caught this — it exercises one build, where the compiler chunk and the entry agree by construction.

A stable path is still the right end state. It needs the compiler built as its own self-contained artifact (its own entry/bundle carrying its own interop helper), or the SPA fallback stopped from answering /assets/* so a rotated chunk returns a real 404 — not a chunkFileNames rename. Recorded in the vite config comment so it is not re-added, and a follow-up ticket will carry it.

The guard keeps its job, and closes two holes

scripts/check-compiler-chunk.mjs is rewritten around the hashed name. It still catches what it was written for — the manualChunks variant that pulled getDefaultExportFromCjs in and made the 2.3 MB chunk statically imported and modulepreloaded — and the same review found two gaps, both closed:

• it now also matches a bare import"./babel-<hash>.js", the form Rollup emits when the importer uses none of the chunk's exports. That shape was eager while matching no from pattern and passing the dynamic-import count. Verified red by injecting one into a built chunk.
• it prints the compiler chunk's own count of content-hashed dependencies — 1 hashed dependency/ies (index-<hash>.js) — must be 0 before it can be renamed. Reported rather than failed: it is the tripwire for when the rename becomes safe.

Still wired into ci.yml's authoring job and master.yml's deploy build.

Verification

	result
pnpm test	847 pass / 0 fail / 2 todo
pnpm typecheck	clean
node scripts/check-compiler-chunk.mjs	ok: assets/babel-U33ShAoN.js, lazily imported by index-DASxLQnA.js only; 1 hashed dependency/ies
same, with a bare import injected	red — imports babel-U33ShAoN.js statically — the chunk is no longer lazy
pnpm e2e e2e/preview-recovery.spec.ts	4 passed / 5 live-skipped

The e2e route glob goes back to **/assets/babel-*.js*. It is not vacuous after the change: the test asserts the recorded request list equals ["(bare)", "?hotRetry=1"] before Restart, which only holds if the route fired.

One review note recorded but not acted on here, since it predates both PRs: playwright.config.ts's webServer swallows --port/--strictPort (pnpm eats them), and with reuseExistingServer: !CI a local run can silently attach to another worktree's server on 4173 and test a stale build. It bit both the review and me; both runs above used an own-port preview instead.

🤖 Generated with Claude Code

---

Note

^{Cursor Bugbot is generating a summary for commit c1d64c9. Configure here.}