docs: settle whether Turbopack traces without standalone — it does

[jouwdan]

Answers MEI-138, the question MEI-134 left open. Documentation only — nothing needed reverting either way, as the ticket predicted.

The answer: yes, and the include globs are applied too

MEI-134 established that outputFileTracingIncludes is honoured, but every measurement came from output: 'standalone'. next.config.mjs skips standalone when VERCEL is set, and which configurations Turbopack emits trace files for is decided inside the Rust binary — not readable out of next/dist.

Two builds of apps/community, differing in nothing but that setting:

	output unset (VERCEL=1)	output: 'standalone'
standalone/ tree	absent	present
.nft.json emitted (excluding the standalone tree's copies)	136, of which 132 routes	136, of which 132 routes
a route's .nft.json (covered by '/**')	2958 files, 1318 @swc/helpers, 324 under esm/, 3 LICENSE	identical
routes carrying no esm/ entry	0 of 132	0 of 132
next-server.js.nft.json (not covered by '/**')	5 @swc/helpers, none under esm/	5 @swc/helpers, none under esm/

The route figures reproduce MEI-134's numbers exactly (1318 / 324), which independently corroborates that measurement. So the glob is load-bearing on every route this app builds, not only on the Docker and self-hosted paths — the second of the ticket's two branches ("the glob is inert on Vercel") is ruled out.

Two traps worth recording, both of which I hit

The ticket's own commands would have produced a wrong reading, so the docs now warn about both:

1. grep -c is misleading here. Each .nft.json is one long line, so grep -c 'swc/helpers/esm' reports 1 for a file holding 1318 matches. It looked like the includes were not being applied until I parsed the JSON. Real counts need json.load/jq.
2. next-server.js.nft.json is at the dist root, not under server/. Looking for it under server/ reports it absent in both builds, which reads like a difference between them and is not one. A standalone build also leaves a second copy of every route's file inside standalone/, which inflates a naive find | wc -l by 132.

What this does not establish

That Vercel's own builder assembles a function from those files. That is the documented mechanism and the reason to expect it, but it was not measured and cannot be from a local build. The docs say so explicitly rather than implying the stronger claim.

The narrow consequence, per the ticket: a packages/db/migrations/**/* entry was worth trying for the missing migration files on Vercel rather than dismissed as inert. vercel.md gains a clause saying so — while keeping MEI-133's schema check as the better design, for the lock-contention and function-timeout reasons that bullet already gives.

Changes

File	Change
docs/contributing/development.md	New paragraph beside the glob with the table above, the reproduction command, and the two measurement traps. Also corrects MEI-134's "three" @swc/helpers entries in next-server.js.nft.json — it measures five now.
docs/getting-started/deployment/vercel.md	The installer bullet's "nothing lists them in outputFileTracingIncludes either" was accurate but incomplete; it now notes that listing them would not have been inert, so the schema check reads as a choice rather than the only option.

Both probe builds (.next-probe, .next-probe-standalone) were deleted, as the ticket instructed — git status is clean apart from the two documents.

Validation

• pnpm verify passes (exit 0) — 467 test files, 8189 tests, lint, guards, all three typechecks.
• pnpm docs:links:check and pnpm site:docs:check pass, so the new cross-links and the #building-where-vercel-looks anchor resolve.
• The two probe builds themselves both exited 0.

---

Generated by Claude Code