ci: run the slowest test packages first - #29457

go test starts packages in argument order, and ./... expands alphabetically, so the heaviest packages started last and set the end of the run. In the baseline test-go-pg (ubuntu-latest) job (run 35182300886, 423s test step) enterprise/coderd (110s) did not start until t=313s.

scripts/test_packages.sh expands the TEST_PACKAGES patterns with go list and prints the five 100s+ packages first, then the light packages (their builds and runs fill the remaining slots instead of forming a tail), then the 20-100s packages. make test, make test-race and make test-timings pass the ordered list to gotestsum. No test code changes; the CI action is unchanged because it already goes through make test. The static list only needs to be roughly right: packages missing from the pattern set are skipped and everything else keeps go list order (verified with the macOS/Windows pattern ./cli/... ./enterprise/cli/... ./agent/...).

Ordering is confirmed locally with go test -json -p 1 ./scripts/release ./agent/agentchat ./tailnet/proto: packages start in argument order.

Before / After

PR CI normally runs with the Go test cache (only main sets TEST_COUNT=1), so the numbers below come from three runs on this branch with test-count: "1" temporarily forced (that override is reverted in the final commit):

The three orderings are within noise of each other and about 10% under main. The job is close to saturated: starting eight heavy packages at once inflates each of them (coderd 224s to 271s), which is why the five-first order was kept: it leaves coderd, the critical path once the fixture PRs land, least inflated (238s). Further wall-time gains come from the fixture and clock PRs in this series, not from ordering.

Validation

• shellcheck scripts/test_packages.sh
• ./scripts/test_packages.sh lists the same 373 packages as go list ./...; the macOS/Windows pattern yields 45 with cli and enterprise/cli first
• make test TEST_PACKAGES='./coderd/httpapi/... ./scripts/release' TEST_COUNT=1 passes
• test-go-pg passed on ubuntu, macOS (BSD awk needed newline-free -v values, fixed), Windows and PG 17 on the measurement commits
• Pre-commit hook (full make pre-commit, since the Makefile changed) passed

Part of the test-suite speedup series: #29458 (mock clocks), #29459 (notifications golden), #29460 (migrations chain), #29461 (dead code), #29462 (server tailnet DERP poll fix). All independent.

Remote validation

Remote run on a dogfood workspace (separate compute), chat https://dogfood.cdr.dev/agents/5bbe73d0-866a-4841-9cc3-c5e5b4d97e44, tested head ee5d3be95da (and the earlier 37754184d0f): ./scripts/test_packages.sh lists 373 packages, identical set to go list ./..., no duplicates, five heavy packages first; make test TEST_PACKAGES='./coderd/httpapi/... ./scripts/release' TEST_COUNT=1 ran exactly those two packages. Endorsed verdict: PASS.

Xum (an AI agent) authored this PR on behalf of @ibetitsmike.