fix(build): one builder per package, so the required gates stop racing

[blove]

The workspace graph was declared twice: once in dependencies, and again by hand inside the script strings. pnpm already orders -r runs topologically, so the second copy bought nothing — and it let two processes build the same package at the same time.

The overlap, as measured

At the base I measured (fca7a789), @pretable/core and @pretable-internal/renderer-dom were topological siblings — both depended on grid-core, neither on the other — so pnpm -r started them at the same instant. A cold pnpm -r --filter './packages/*' build shows it directly:

20:33:43  grid-core build$ tsc -b
20:33:45  grid-core build: Done            <- pnpm already built it, topologically
20:33:45  core build$ pnpm --filter grid-core build && tsup
20:33:45  renderer-dom build$ pnpm --filter text-core build && ... layout-core ...
20:33:46  renderer-dom build: > layout-core > tsc -b     <- concurrent with core's

Two tsc -b processes emitting one dist/ is unsafe by construction: writeFileSync truncates before it writes, so a concurrent reader can see layout-core/dist/index.d.ts empty and report "is not a module".

I could not catch the corrupted read on this machine (0 failures in 10 cold races — APFS writes a 1 KB .d.ts too fast). So I measured the precondition instead. Two cold tsc -b runs launched together emitted that file twice in 2 of 3 runs, once in the third. That split is the race: whether the second process's up-to-date check lands before or after the first process's write is a coin flip, and on a 2-core runner with a cold page cache the window is wider than here.

Correction, after rebasing

That particular sibling pair no longer exists. #448 landed while this branch was open and made renderer-dom depend on @pretable/core, which linearizes the tail of the graph:

LEVEL 2: @pretable/core
LEVEL 3: @pretable-internal/renderer-dom
LEVEL 4: @pretable/react

Levels 2–4 are now single-occupancy, and the two packages that share level 1 (grid-core, bench-runner) have disjoint project references. So on today's main the race is dormant, not fixed — it was real and measured at the commit I measured it on, and what made it possible is still in the tree. It comes back the moment a second package joins any level, which is an ordinary thing to do and carries no warning that it re-arms a flake in a required gate.

I'd rather say that plainly than leave the PR claiming a live failure.

The change

Each package builds only itself. The root test and typecheck build the packages once, up front, before running either.

Apps keep their standalone prepare:deps — pnpm --filter @pretable/app-bench dev still works from an empty tree — but derive the closure with pnpm's <pkg>^... filter instead of listing siblings, so it stays correct when the graph changes. (It already has: this branch needed no edit to track #448's rewiring, while the hand-written chain it replaced named @pretable/react and would have kept working only by luck.) The two CI jobs that build only the published packages now ask for their dependency closures (<pkg>...).

Wall clock locally is unchanged (8s both ways) — the redundant builds were mostly up-to-date no-ops here. This is a correctness fix, not a speed one.

The guard

scripts/__tests__/workspace-scripts-own-one-package.test.mjs holds the invariant: a workspace script builds its own package and nothing else. It discovers the workspace rather than listing it, so a package added tomorrow is covered with no edit.

It is mutation-tested. Re-adding the exact prefix this PR removes fails it; so does a filter naming a package that does not exist; and the sanctioned derived form still passes. Its first draft passed both mutations — it skipped any --filter argument containing /, which is every scoped package name, so it was checking nothing. That blind spot is now called out in the code so the next person doesn't reintroduce it.

Verification

Cold, rebased onto cf3ff2b4: typecheck, test (react 1227, website 556, bench 156, renderer-dom 127), build including the Next site, lint, format, api:check, lint:packaging, typecheck:public, typecheck:performance, the rewritten CI build command, and a standalone prepare:deps from an empty tree — all exit 0.

One note for reviewers pulling this branch: #448 added a workspace dependency, so a node_modules predating it will fail renderer-dom's build with Cannot find module '@pretable/core'. That is a stale install, not this change — pnpm install clears it.

No changeset: this changes how the workspace builds, not what any published package does.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
pretable	Ignored		Aug 17, 2026 4:18am

[github-actions]

Vercel preview ready

Preview: https://pretable-s4l0sopmb-cacheplane.vercel.app
Commit: a4b8c950213c6ecbeced46836bd7742870c12122

_{Updated automatically by the deploy-preview job.}