[maxigimenez]

Follow-up to #50. The 0.0.4 publish failed at Test.

What broke

publish-cli.yml runs Lint → Test → Build. ci.yml builds first — which is why #50's PR checks were green and the publish was not.

The bundled-manifest test added in #50 read packages/common/dist/ to check that every exports entry names a file the build emitted. On the publish path there is no build yet, so all fourteen entries looked unbuilt:

+ "  @sentinel0/common/github → ./dist/github.js (not built)",

The fix

The test no longer depends on a build. It maps each ./dist/<module>.{js,d.ts} target back to src/<module>.ts and asserts the source exists, and flags any target shape the mapping doesn't recognise so it can't pass by accident.

Guarding it with skipIf(!existsSync(dist)) — the pattern common's circular-import guard uses — would have been the worse fix: green, and testing nothing, on the one path that publishes. prepare-package.mjs still asserts against real dist/ at pack time, where a build has just run and cannot be absent.

publish-cli.yml builds before testing anyway, matching ci.yml. That also un-skips common/test/built-output.test.ts, which imports the compiled package to catch circular imports the source alias hides — it had been skipping silently on every publish.

Verification

From a clean tree (pnpm clean), both orders:

	no dist/	after build
CLI suite	57 passed	57 passed
common suite	3 passed, 1 skipped	4 passed

And the test is not vacuous — adding an ./ghost export with no source module fails it with no dist/ present:

× names only modules packages/common actually has
+   "./dist/ghost.js — no src/ghost.ts to build it from",

Then the full publish sequence in the new order — lint, build, test, exec-bit, pack — followed by a clean npm install of the tarball: the bundle gate passes and sentinel0 start reaches its config check.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EkUWsM92gcbit4p6hEhuyV