Move the test image pins out of test bodies and into testutil

[Peeja]

Answers your review question on #17. Stacked on it, because it moves the exact lines #17 touches — unstacked, whichever merged second would conflict.

Two new files, four one-line call-site changes. No git subtree add, so rule 8 does not apply.

You were right, and the repo already agreed with you

The strings pre-date #17 — it only appended @sha256:… to literals already inline, so the placement was upstream's. But the instinct holds, and piri/pkg/internal/testutil/minio.go is this repository's own better answer: a named const, a doc comment saying why the pin is what it is, and an environment override for pointing at a local build.

indexing-service/pkg/internal/testutil/images.go   MinioImage(), ValkeyImage()
swarf/internal/testutil/images.go                  PostgresImage()

hilt and sprue are halfway there — raw literal, but in a testutil package. The four sites here were the least-good shape in the repository and are now the best one.

Why exactly four, rather than a line I drew

I surveyed all 19 pinned image references in Go across six modules before deciding:

where	count	verdict
already in a testutil package	7	leave — hilt, piri ×3, sprue ×2
smelt production code	4	leave — smelt generates compose files; the image belongs there
smelt build_test.go heredocs	3	leave — FROM alpine:… is test input, not an image the test pulls
ingot/itest awsCLIImage	1	leave — already a named const, with five lines on why the CLI version is the surface under test. Better than anything I'd impose
inline literal passed to a Run call	4	moved

So "scope it to the duplication rather than these four lines" turned out to be the same four lines. I said that differently before checking; the survey is what settled it.

Two things I deliberately did not do

No shared cross-module constant. postgres:16-alpine@sha256:cf78e766… is named in eight places across five modules (I told you six earlier — the survey says eight). A single source of truth needs a shared package, which across five separate Go modules means a new in-repo module plus five replace edges, to carry a test constant. That cure is worse than the disease. Renovate groups them instead — once it is installed.

No guard, for the reason #17 already gives. "No image literal in a _test.go" is clean and list-free, but it can only see references that are already pinned — a 64-hex digest is unambiguous, an unpinned image is not. It would enforce placement while reading as though it enforced pinning. That is exactly the partial-guard shape this consolidation keeps deleting.

One thing left on the table

swarf's two call sites are otherwise identical — the same twelve lines of database, user, password and wait strategy, twice. A RunPostgres(t) helper would remove real duplication, but that is a different refactor with a different justification, and it is not what you asked about. Say the word.

Verified

go vet clean on indexing-service and swarf; gofmt -s -l . reports 0 files repo-wide; both existing guards still pass (26 base images, 15 compose images); and no _test.go in the tree holds an image literal any more except ingot's documented const.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CGAGAib517Ae1kg8SCdcEt

---

Generated by Claude Code

[Peeja]

LGTM; merging when green.