Lint policy

CI runs golangci-lint with a narrow, high-signal set so Pulumi-heavy graphs stay tractable. Expand deliberately; do not dump every available linter into CI.

Enabled (required)

Linter	Role
govet	Compiler-adjacent correctness
staticcheck	Bug-finding (selected checks; see .golangci.yml)
ineffassign	Ineffectual assignments
misspell	Identifier/comment typos
unused	Dead code
errcheck	Unchecked errors
revive	Style subset (narrow rules in .golangci.yml)
gocritic	Selected diagnostics (#diagnostic only; opinionated off)

govulncheck runs as a separate CI step inside go-verify (not inside golangci-lint).

CI partitioning (QUALITY-06)

Lint no longer runs as one golangci-lint run ./... job. The import graph forces six partitions in .github/workflows/ci.yml:

Partition	Packages	Why separate
aws	./internal/cloud/aws/...	Largest Pulumi AWS SDK graph
gcp	./internal/cloud/gcp/...	Independent GCP SDK graph
ovh	./internal/cloud/ovh/...	Independent OVH SDK graph
scaleway	./internal/cloud/scaleway/...	Independent Scaleway SDK graph
core	Explicit list of everything else (see workflow)	Go patterns cannot subtract ./internal/cloud/... from ./internal/...
aggregate	./cmd/magelift/...	Only binary that imports all four providers — irreducible

Ordered fallbacks when a partition OOMs or the runner is reaped (do not raise timeouts as the first response — a dead runner with no lint output is a memory signal, not a slow one):

1. Keep cache-prime as module-download only (compiling the full module was SIGTERM'd on private runners).
2. Serialize the lint matrix (max-parallel: 1) and run go-verify after lint.
3. Soft-fail cache-prime so a cold matrix still gates merges.
4. If a single partition still dies, shrink that partition's package list further before touching timeouts.

Coverage guard: go test ./internal/lintcoverage/ asserts the union of matrix patterns equals go list ./... (minus .golangci.yml path exclusions). A new top-level package that joins the module without joining a partition fails that test.

Force-all verification run

Path filters skip php, docs, and workflow-lint on Go-only changes. To reproduce a full make verify mapping in one CI execution, dispatch with the all input:

gh workflow run ci.yml --ref main -f all=true

Map jobs → verify targets: generate-check / cli-docs-check / fmt-check / test / license-check in go-verify; php-test in php; docs in docs; workflow-check in workflow-lint.

force-all proof: Hosted Actions minutes are available on the public repo. Record the latest force-all URL and per-target outcomes below when a workflow_dispatch with all=true completes.

gh workflow run ci.yml --ref main -f all=true

Map jobs → verify targets: generate-check / cli-docs-check / fmt-check / test / license-check in go-verify; php-test in php; docs in docs; workflow-check in workflow-lint.

Run	Date	Outcome
30836267745	2026-08-03	force-all green (lint, go-verify, images, varnish, floci, php, docs)

go test -race measurement

Criterion 1 requires go test -race ./... inside go-verify. -race roughly doubles per-package memory over a graph that peaks near 8.5 GB when provider SDKs link into one binary, so fit on the 2 vCPU / 7 GB runner class must be measured, not assumed.

Environment	Command	Wall time	Peak RSS	Outcome
Local (serial sample)	GOMAXPROCS=1 GOFLAGS=-p=1 go test -race ./internal/lintcoverage/ ./internal/usererr/ -count=1	~4.7s	~102 MiB	pass (narrow sample only)
Local (full race)	GOMAXPROCS=1 GOFLAGS=-p=1 go test -race ./... -count=1	deferred-local	deferred-local	Prefer CI; see agents/skills/magelift-serial-builds for local smoke guidance
CI go-verify (2 vCPU / 7 GB)	go test -race ./...	~12m	n/a	pass on run 30836267745

If CI fails on memory or the runner is reaped with no test output, partition the test step by the same six package groups as the lint matrix — do not raise timeout-minutes as the first fix. Record the partitioning reason here when that happens.

Deferred linters

Linter	Why deferred
gocyclo / cyclop	Pulumi composition graphs trip low thresholds; revisit with a high bar
gosec	Prefer govulncheck + CodeQL + Trivy for security signal
wrapcheck	Noisy across CLI/fmt.Errorf boundaries

Suppressions

Prefer fixing the finding. If a false positive is unavoidable, use a scoped //nolint:<linter> with a short why on the same line. Package-wide nolint is not allowed without an ADR-linked rationale.