test: run the runes-module tests through the compiler that owns runes

[PathGao]

114 test files carry 678 source-shape assertions and 88 of them import scripts/sourceTree.ts, which reads src/ as text. The reason is in that file's own header: node --test cannot import .svelte. Extracting more modules does not move that number — the last round of refactoring moved it by 5.

So this changes the runner, for one specific class of test, as a pilot.

The class

.svelte.ts runes modules — stores, sessions, tab identity, save guards. Real logic that currently cannot be executed at all, so the tests hand-write a $state proxy shim onto globalThis and hope it matches the compiler.

Three files migrated: reopenDirtyDocument, tabPathIdentity, settingsPersistence. They keep their names but move to *.spec.ts, which the existing scripts/*.test.ts glob does not match — two runners, zero overlap, either can be stopped.

What the pilot had to prove

1. The shims come out — 84 lines. No .spec.ts installs $state / $derived / $effect onto globalThis any more. The stores compile through the Svelte plugin; jsdom supplies window / localStorage / navigator for real. settingsPersistence lost ~62 of the 84 (full rune shim plus localStorageShim and windowShim).

2. $effect really flushes. settingsPersistence.spec.ts now constructs a real SettingsStore — whose constructor installs ~30 effects inside $effect.root — calls flushSync(), mutates one field, flushes again, and diffs actual localStorage. It asserts the changed-key set is exactly ['editor.minimap'], then exactly ['editor.fontSize','editor.language'], then [] for a no-op.

The old version hand-simulated the flush loop and then asserted on its own simulation, so it could not fail. The registeredEffects.length === entryCount + 1 counting test is gone with it — subsumed, because the behaviour is now observed rather than counted.

Negative control: setting resolve.conditions: [] makes exactly 5 of 54 fail — precisely the $effect-dependent ones. That is simultaneously proof the trap is real and proof the effects are running.

3. CI does not get slower. npm test 22.3s → 18.2s (three files left the glob). npm run test:vitest 2.8s. Combined 21.0s, marginally under the baseline. vitest's cost is per-process, not per-file, so further migrations add much less than linearly.

The shims had already drifted, and nobody knew

createRecordingStore() forwarded with Reflect.get(target, prop, receiver). The real compiler backs every $state with a #private field plus a getter, and invoking that getter with the Proxy as this throws Cannot read private member #minimap. Under the identity shim these were plain properties, so it passed.

This is the concrete form of the concern: a shim passing is not the same as the code passing.

Other potholes, for the next person

• import.meta.url is http:// under vitest (Vite serves the test files), so readSource(new URL(…, import.meta.url)) dies with The URL must be of scheme file. The cwd-relative string form, which readSource already documents, works. Noted in its doc comment.
• jsdom does not fire storage for same-document writes (correct per spec). Tests now dispatch a real StorageEvent, which exercises the store's actual addEventListener instead of a captured callback array.
• vite.config.js exports an async factory, so mergeConfig must call it first. It merges cleanly with sveltekit() included — no need to drop to a bare svelte() plugin.
• vitest.config.ts was invisible to npm run check (tsconfig include listed vite.config.* only). Adding it immediately caught a wrong UserConfig import path.

The boundary is written down, with reasons

AGENTS.md gains a rewritten Testing section: *.spec.ts when the subject is a .svelte.ts runes module or needs a real DOM; *.test.ts otherwise, because moving a passing file buys nothing; do not migrate .svelte component tests, because jsdom standing in for Monaco/mermaid/KaTeX is a swamp and "the handler is wired up" is weaker than what it would replace.

Then a never-migrate section that gives the reasoning rather than just the list:

• Cross-language contracts — the only thing connecting Rust to TS is the spelling; running the TS proves the TS.
• Single-implementation conventions — these are claims about the absence of a second implementation, and executing the one that exists cannot observe a second.

That is what makes the two-runner split a defensible end state rather than an unfinished migration.

Known before rolling out further

• SettingsStore discards the disposer $effect.root returns, so every store a test constructs leaks live effects into that file's environment. It did not bite here, but a file that constructs stores and flushes repeatedly could see cross-test writes. The fix belongs in src/ and should land before file Border Around Window on Windows 10 #10.
• foldStatePerDocument.test.ts is not in this class. It slices functions out of a .svelte file with a hand-written brace matcher; a runner swap does not fix that. The subject needs extracting into a .svelte.ts module first.

25 files still install rune shims. Each was ~15 minutes once the config was right, and the API delta is one import line.

Verification

• npm test 953 pass / 0 fail
• npm run test:vitest 54 pass / 0 fail
• npm run check 766 files / 0 errors
• No file under src/ or src-tauri/ was touched.

🤖 Generated with Claude Code