One unloadable test file no longer takes down the whole run

[DmitrySharabin]

A test file that cannot be imported — a missing dependency, a syntax error — used to kill the process with a raw unhandled-rejection stack before any test ran, including the healthy files beside it.

Two paths handled it, both wrong, in opposite directions:

• First run. getTestsIn logged Error importing tests from … — but the tagging loop at the end of resolveLocation then re-imported the same file with a bare import(url), and that rejection killed the process anyway. The handler was effectively dead code. The glob branch had no handler at all.
• --watch. rerun() logged to a console the alt screen paints over and kept the stale passing subtree, so a file that broke while you watched it stayed green until the next successful save.

The change

An unloadable file becomes a failing test:

function failedTest (err, name) {
	return { name, run: () => { throw err; } };
}

That is the whole design. The exit code comes from the existing stats.fail > 0 ? 1 : 0 line, the error renders in the tree where interactive mode can actually show it, and --watch recovery works because fixing the file is just another re-run. No new state, no new exit-code plumbing, no console channel that the alt screen hides.

importTests(url, base) replaces the two near-duplicate import sites, so the load-failure policy has one owner.

(All tests) ❌ 1/2 PASS 1/2 FAIL          exit 1
 ├── good.js ✅ 1/1 PASS
 └──  FAIL  broken.js: Got error ERR_MODULE_NOT_FOUND: Cannot find module './nope.js'…

Verification

122/122. Both new assertions go red against the unfixed runner, and the plausible wrong fix — log the error, drop the file — is caught too: it yields 1/1 PASS and exit 0.

Exit 1 and a correct broken.js label for a directory, dir/*.js, and a bare *.js. --watch checked expanded in tmux across five transitions: broken on first run → fixed → re-broken → broken again while broken → deleted.

The tests deliberately spawn from run() rather than beforeAll(): a throwing beforeAll degrades its children to SKIP, and a skipped test cannot fail CI. With an injected fixture fault they now report 2/2 FAIL and exit 1 instead of a green 13/13 skipped.

Diff stats

Substantive lines only (blank- and comment-only changes excluded).

File	Added	Removed
src/env/node.js	28	31
tests/run.js	42	0

Not covered

The rerun() half is verified by hand, not by the suite — there is no pty/watch harness, and rerun has no seam (module-private, needs currentRoot, isInteractive, watchers and debounce timing). A refactor of rerun could reintroduce the stale-green bug with the suite still green.

Out of scope

• A CJS file (no "type": "module") with a syntax error still hard-crashes: Node's module-detection retry rejects a promise hTest never receives. Unchanged from main.
• **/*.js yields base = "**", so labels read ../broken.js. Pre-existing path.dirname(location) behavior, now also reaching failing-test names.
• A test that throws a non-Error crashes the runner the same way — same family of bug, different cause. Filed as Throwing a non-Error crashes the runner and takes down the whole run #183.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RaXmFiBg3UGUMuGFXjn5Ki

[netlify]

✅ Deploy Preview for h-test ready!

Name	Link
🔨 Latest commit	777ed38
🔍 Latest deploy log	https://app.netlify.com/projects/h-test/deploys/6aa3f5a61f489a00088f1a1a
😎 Deploy Preview	https://deploy-preview-184--h-test.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.
🤖 Make changes	Run an agent on this branch

---

To edit notification comments on pull requests, go to your Netlify project configuration.