fix(exports): correct ./components/* source targets for two non-tsx modules

[YevheniiKotyrlo]

Problem

Fixes the four broken targets reported in #445.

./components/* declares source and react-native as ./src/components/*.tsx, but two modules the pattern matches are not authored .tsx:

$ ls src/components/ | grep -E '^(index|copyComponentProperties)\.'
copyComponentProperties.ts
index.cts
index.ts

Export conditions match on key presence, not target existence, so there is no fallthrough to import/require — both subpaths are a hard ERR_MODULE_NOT_FOUND under either condition:

$ node --conditions=source -e "…require.resolve(…)"
[source] react-native-css/components/copyComponentProperties -> FAILED: MODULE_NOT_FOUND
[source] react-native-css/components/index                   -> FAILED: MODULE_NOT_FOUND

react-native is not dormant — Expo sets it for iOS and Android via unstable_conditionsByPlatform. source is the condition react-native-builder-bob's example-app workflow runs on.

Solution

Two exact keys. An exact key beats a pattern regardless of declaration order, mirroring the existing ./components/react-native-gesture-handler entry.

index takes .cts rather than .ts because that is what the exact ./components entry already declares — the wildcard's .tsx guess contradicted the map's own neighbouring entry.

Tests

src/__tests__/package/exports-source-targets.test.ts — 37 cases, no build needed. Resolution is decidable from source alone, so it walks the exports map the way node does (exact key first, then the pattern whose literal prefix before * is longest) and asserts every source and react-native target names a file on disk.

Two censuses drive it: every literal subpath, and every module in src/components/ reduced to the subpath a consumer would import. A case asserts both are non-empty, so a narrowed map cannot quietly produce zero cases.

Reverting the manifest diff reddens exactly ./components/copyComponentProperties and ./components/index, leaves the other 33 green, and names ./src/components/copyComponentProperties.tsx in the failure — the file the wildcard points at and that is not there.

Verification

Applied against the published 3.0.7 tarball, both subpaths resolve under source and react-native, and every other entry is unchanged. I checked the whole map rather than just these two: 509 declared targets across every entry and every condition, 0 missing after the fix.

That sweep is how the defect surfaced. It walks the real dist/module/** tree, expands each wildcard over it (skipping stems that have an exact key, since exact beats pattern) and asserts every declared target exists. The test above is its source-only half — the part that needs no prepare and so can run in CI; the dist/** half would belong post-build if you want it.

yarn test — 1085 passed, 21 skipped, 3 failed; the three are the Windows-only babel path-separator failures #390 fixes, green on CI. yarn typecheck and yarn lint clean.

Base

This may already be fixed upstream. Current main already declares both ./components/copyComponentProperties and ./components/index with the same targets this diff adds, so the change is a no-op against it. #445 was closed as completed on 2026-09-14. Close it if nothing here is still missing; I will fold whatever is into a fresh one.

Branched off f70c402. main has since taken #451 (a5002c5). 1 of the 1 files this changes also moved there (package.json), and it still merges cleanly onto current main. Every measurement above was taken on f70c402. Say the word and I will re-apply it onto current main.

[YevheniiKotyrlo]

Why no device screenshot, and no test

Both have the same cause: this changes two exports targets, so what it fixes is module resolution — before any frame is painted, and outside anything jest observes.

The measurement is exhaustive rather than illustrative: 509 declared targets across every entry and condition, 0 missing after the fix. Before it, two are a hard ERR_MODULE_NOT_FOUND under either source or react-native, because export conditions match on key presence rather than target existence.

The sweep needs dist/**, so it belongs as a post-prepare check rather than a test. Happy to add it.