test(exports): catch a browser condition that names the node bundle

[sroussey]

Stacked on #717 (claude/wonderful-turing-rjtcnx-ai-types). Assertions only — no manifest is touched.

What breaks

packages/test/src/test/util/ExportTypesPairing.test.ts has six checks and every one of them is internally self-consistent: it compares a branch against itself (does this types name the implementation beside it, is it declared first, does a source entry exist for it, is it shadowed by a sibling). A browser block that is a byte-for-byte copy of the default branch satisfies all six — and routes browser consumers to the node bundle. That is the exact class of bug the file was written to prevent, expressed as a duplicate rather than as a wrong target.

What was NOT changed, and why

providers/chrome-ai and providers/tf-mediapipe carry exactly that duplicated shape today. They are correct as written and are deliberately left alone. Verified before writing any code:

• providers/chrome-ai/src/ contains ai.ts, ai-runtime.ts, ai/ — and no ai.browser.ts;
• providers/tf-mediapipe/src/ likewise.

So ./dist/ai.browser.js is a bundle nothing builds. Repointing either manifest at it would immediately turn this PR's own existing "declares only targets a source entry file can emit" test red. Both are browser-only packages with one entry each; the two bundles genuinely are the same file. Their blocks are inert today, and the new rule fires the instant someone adds the source entry.

The fix

A pure browserSplitViolations(manifest, exportsMap, hasSourceEntry) beside findViolations, with the source-entry probe injected so fixtures can drive both sides of it (the convention unresolvedWorkspaceMessage already uses). Per subpath:

1. take the branch a plain node import resolves (nodeImportTarget — node/import/default only; browser, bun, react-native excluded on purpose, since the point is the target a browser split has to differ from);
2. derive the stem with the existing distStem;
3. probe dirname(manifest)/src/<stem>.browser.{ts,tsx}.

No source entry → skip. That is what keeps chrome-ai green and what keeps the packages/* browser/node split (stem node, no src/node.browser.ts) inert — that layout expresses its split as two peer entries, not a .browser suffix. Entry exists → require a browser condition and that every implementation string inside it is ./dist/<stem>.browser.js. Two distinct messages: no browser condition declared, and browser condition names the node bundle.

Also: orderViolations

It treated only string-valued implementation keys as shadowers, so { "import": { default: … }, "browser": … } reads as two conditions and behaves as one dead branch, unflagged. Replaced with isShadowingImplementation, which also counts an object whose implementation keys resolve unconditionally. A nested object with no unconditional target ({ import: { node: … } }) does not shadow — Node falls through it. String behavior is unchanged, and no existing manifest newly violates (the repo-wide orderViolations assertion stays []).

What the tests catch

• New repo-wide test passing existsSync as the probe, asserting []. It is live, not vacuous: temporarily repointing providers/deepseek's browser target at ./dist/ai.js makes it fail with the expected message (verified, then reverted).
• Fixtures proving it goes red only when the source exists:
  • chrome-ai's exact shape + probe true → the "not the browser bundle" message (the go-red proof; the same fixture is asserted [] against findViolations and orderViolations to show why a separate rule is needed);
  • chrome-ai's exact shape + probe false → [] (today's tree);
  • a correct provider shape (deepseek) + probe true → [];
  • same with the browser key deleted + probe true → the "no browser condition" message;
  • the packages/* browser/node layout + probe false → [] (no false positive).
• Two fixtures for the shadowing predicate: a nested import that always resolves hides a later browser; one that does not, does not.

Verification

• npx vitest run --project test packages/test/src/test/util/ExportTypesPairing.test.ts — 28 passed.
• bun scripts/test.ts util unit vitest — 54 files, 785 passed / 10 skipped.
• npx tsgo -p packages/test/tsconfig.json — clean.

---

Generated by Claude Code

[github-actions]

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	59.72%	37022 / 61992
🔵	Statements	59.23%	38840 / 65573
🔵	Functions	60.74%	7174 / 11810
🔵	Branches	47.89%	18767 / 39184

File Coverage

No changed files found.

Generated in workflow #3029 for commit c6906ca by the Vitest Coverage Report Action