test(electron): make dev-resolution tests self-sufficient on fresh clones

[ozymandiashh]

Closes #681.

Problem

On a fresh clone (no npm run build), the two Vite-development resolveCodeburnPath tests failed with .toMatch() expects to receive a string, but got object: the dev branch only returns the repo CLI path when dist/cli.js exists and is executable.

Fix

Per the issue's preferred direction, the tests now create the fixture they depend on:

• fakeRepoDevCli() writes a stub executable dist/cli.js at the exact path the production sourceDevBin fallback resolves, and snapshot-restores any pre-existing real build artifact (content and mode restored byte-identically in afterEach; the fixture deletes only what it created, and dist/ is kept when it was non-empty before).
• A fail-fast precondition asserts the parent-level dist/cli.js candidate (which resolution prefers under vitest's module layout) does not exist, naming the offending absolute path instead of failing later with a confusing toBe mismatch.
• No skip-when-absent: the tests run and pass with no build present, and still pin the preference order (repo dev CLI over a persisted stale global).

Production code untouched.

Verification

• npx vitest run app/electron/cli.test.ts on a checkout with NO dist/cli.js: 19/19.
• Same suite with a pre-existing sentinel dist/cli.js (distinct bytes, mode 0711): 19/19, then the sentinel is restored byte-identically with its exact mode (cmp clean).
• npx tsc --noEmit clean.

[iamtoruk]

The core fixture is sound and the assertions got stronger (exact-path toBe instead of toMatch, real statSync and X_OK checks still exercised), but three things before this can merge.

1. It is stale against main: there is a content conflict in app/electron/cli.test.ts, and main gained a third fresh-clone failure after this branched (the resolveTarget test where the dev repo CLI beats the bundled CLI). As-is, Closes #681 would close the issue with one of its three symptoms still failing. Rebase and cover that test too.

2. The stub writes the real repo dist/cli.js, which is the published bin and exactly what npm run dev spawns. The snapshot-restore in afterEach is careful, but a SIGKILL or worker crash between the write and the restore leaves a developer's real build replaced by a 20-byte no-op, silently. Since there is no CI test job, this suite runs precisely on developer machines where that hurts. The cleaner shape already has precedent in this file: CODEBURN_PATH_DIRS exists for tests and advanced setups. A matching dev-root override env var lets the whole fixture live in the per-test tmpdir and deletes the snapshot machinery, the precondition check, and the crash window.

3. Please drop the beforeAll repo-scanning guard that globs test files and regex-scans imports. It misses import-star and indirect uses, will spuriously fail the suite when someone adds a legitimate isolated resolver test, and likely self-detects on Windows path separators. It is the needless-defensiveness pattern we strip in review.

Happy to re-review quickly, the direction is right.

[ozymandiashh]

Reworked per review, force-pushed as a fresh implementation on current main:

• Env-var shape adopted: CODEBURN_DEV_REPO_ROOT on the Vite-dev resolution branch (mirroring the CODEBURN_PATH_DIRS precedent). The whole fixture now lives in a per-test tmpdir; the real repo dist/cli.js is never written, so the crash-window concern is gone along with the snapshot-restore machinery and the precondition check. With the env unset, resolution is byte-identical to before (both candidates, same order), covered both ways.
• All three fresh-clone failures covered on current main, including the newer resolveTarget dev-repo-beats-bundled case; verified on a checkout with no dist/cli.js at any point (27/27).
• beforeAll scanning guard dropped.