Repository navigation
Fence: go.sum completeness across the multi-module tree (nobody's beat today — #975 class) #976
Description
#975 exposed the class: a merge can leave importers' go.sum incomplete and main still passes every gate that runs, because go.sum completeness is currently nobody's beat — check-deps-sync.py checks go.mod Require sync only, CI is frozen, and per-PR verification runs the PR's touched modules, not every importer of a changed dependency surface.
Fence shape (proposal): extend the deps checker (or the local CI gate #955 is building) with a tree-wide go build ./… per module directory — any module dir failing to build against its own go.sum is red. Cheap (builds only), deterministic, and it is exactly the #975 repro.
Acceptance: a deliberately-stripped go.sum entry in any module makes the fence fail naming the module; the current tree passes.
Activity
Amperstrand commentedon Oct 17, 2026
Fence delivered — PR #985:
tests/contract/check-gosum-completeness.shprobes every module under src/ withgo build ./...from inside the module dir (GOWORK=off, -mod=readonly), ~2s warm for all 18, fails naming modules + quoting the missing-entry lines. Both negative controls proven: the hermetic--selftest(graph-derived victim dependency — the naive last-line strip was tried and did NOT fail, so the fixture derives its victim fromgo list -depsto stay rot-proof), and the live one — on current main it is RED on exactly merchant_types / upstream_detector / upstream_session_manager (the #975 breakage), per this mission's battery rule. Sequencing for the owner: land #975 first → this PR's lanes go green; either order proves the fence. Wired into release-check's deps leg + both test.yml twins' deps-and-imports (twins have pre-existing drift on main — noted in the PR, not touched). No merges performed.fence: go.sum completeness across the multi-module tree (#976) (#985)