Ask the toolchain which files are ours, in one place

[verygoodsoftwarenotvirus]

The problem

The tools that walk the filesystem each decided for themselves which Go files belong to this module, and they disagreed three ways:

Where	How it excluded vendor
scripts/format_golang.sh	find … -not -path '*/vendor/*'
scripts/format_imports.sh	find … -not -path '*/vendor/*'
.github/workflows/formatting.yaml	gofmt -l . | grep -Ev '^vendor\/'
scripts/goimports.sh	it didn't — go tool goimports -w .

That last row is a bug rather than a duplication. goimports -w . walks from the working directory, so with a vendor tree present make format rewrote vendored third-party source. Measured against this repo's own dependencies: 171 of 2681 vendored files reformatted before I interrupted it, and the run was still going after six minutes.

The fix

scripts/go_files.sh answers the question once, by asking the Go toolchain:

go list -e -f '{{.Dir}}' ./...   # then find -maxdepth 1 over those directories

./... does not descend into vendor/, testdata/, or any _- or .-prefixed directory — which is exactly why every wildcard-driven target here (test, lint, go fix, fieldalignment, tagalign) has never needed an exclusion at all. Only the filesystem walkers ever did, and now they inherit the same answer instead of each maintaining their own.

-e keeps a package that does not compile in the list, because formatting a file is most useful exactly when it is still broken. -maxdepth 1 because what go list names are package directories, and a package is entitled to a testdata directory of its own.

Two details worth reviewing closely

Both were found while testing this, and both are the sort of thing that hides:

It fails loudly rather than emitting an empty list. An out-of-sync vendor/modules.txt makes go list exit non-zero — not exotic, it's what you get from editing go.mod without re-vendoring. The first version of this returned an empty list in that case, and every formatter silently formatted nothing while the CI check reported "clean". Silence read as success.

Callers read it through a file, not < <(go_files.sh). Process substitution discards the exit status of what it runs, which is precisely how that empty list went unnoticed. The workflow sets pipefail for the same reason: xargs exits 0 on no input and would otherwise mask the failure.

Verification

Against a real go mod vendor of this module's dependencies (2681 vendored Go files):

Check	Before	After
Vendored files rewritten by make format	171+ (interrupted)	0
make format wall clock	>6 min, unfinished	19s
make format on a clean tree	—	no-op
CI formatting check	—	clean
Every path when go list fails	silent success	exits non-zero

shellcheck passes on all of scripts/, and go test -race ./... is unchanged.

Provenance

This came out of primandproper/tarpaulin, which had drifted its own copies of the same exclusion (*/vendor/* in the scripts, ./vendor/* in the workflow) and has the equivalent change applied. Nothing here is tarpaulin-specific.

[verygoodsoftwarenotvirus]

LGTM