fix(install): check the schema instead of migrating in the function

[jouwdan]

Third bug from the same deploy, after #240 and #243. The board now renders its own text and its own styling; /install step 1 fails:

"Apply migrations" did not finish. Cannot find the migrations. Set MIGRATIONS_DIR to the folder holding meta/_journal.json — looked in /var/task/node_modules/@meith/db/migrations, …

Cause

@meith/db ships migrations in its files allowlist, and the package's JS is plainly in the function — the code ran far enough to raise this. The .sql files and meta/_journal.json are not, because they are data nothing imports.

Nothing can put them there either. next build defaults to Turbopack in Next 16, and Next skips trace collection entirely for it — next/dist/build/index.js:1627:

if (bundler !== _bundler.Bundler.Turbopack && !isGenerateMode && !buildTracesPromise) {
    buildTracesPromise = collectBuildTraces(...)
}

So outputFileTracingIncludes never runs. That also makes the @swc/helpers entry from MEI-120 inert — a second guard believed to be protecting the build that does not execute. Filed on MEI-133 for a separate look; not touched here.

Why the step had nothing to do anyway

vercel.json's build command is community migrate && forum-web build --at-root, which migrates in the build container where the files do exist. The schema was current before the app ever served a request.

The fix

The step confirms every table is there and stops with the names of any that are not. The names come from the schema definitions — ordinary imported code, present wherever the board runs at all — so nothing reads a file. 72 tables are discoverable that way.

A serverless function was the wrong place for this regardless of the missing files: several cold starts contend for one migration lock, and the function timeout bounds how long a migration is allowed to take. The board migrates where it deploys — community migrate in the build command, docker-entrypoint.sh before a container serves anything — so self-hosted behaviour is unchanged.

Two things found on the way

• The failure copy was being built in @meith/db, which has no business holding user-facing prose and no translator to hold it with. It moved to the catalog.
• install.raced.sealed and install.raced.inFlight were already thrown as catalog keys and rendered raw — an operator hitting either saw the literal string install.raced.sealed. Rendering the step error through fromCopy fixes that too; it falls back to the raw string, so every other thrown message is unchanged.

Verification

• pnpm verify green — 7,876 tests, 449 files
• the schema check is exercised with tables renamed away, which is the path CI could not reach before: its boards install a full node_modules, so the migration files were always present
• a test holds the form's copy to the keys the server throws, so a key added without copy behind it fails rather than shipping as raw text — confirmed by removing the new key and watching it fail
• no code comments; the reasoning is in docs/vercel.md and docs/quickstart.md, per AGENTS.md

Fixes MEI-133.

---

Generated by Claude Code