fix(sbom): regenerate the SBOM when the GOST configuration changes

[reyreavman]

Summary

Changing build.sbom.gost did not regenerate the attached SBOM: the build reused the artifact from the previous run, so the document kept the old GOST properties. Reproduces from a clean checkout on any image whose SBOM has no base or imported document:

project: gost-repro
configVersion: 1
build:
  sbom:
    enable: true
    standard: "cyclonedx@1.6"
    gost:
      attackSurface: yes
      securityFunction: yes
---
image: app
from: registry.werf.io/werf/scratch

Build with --repo, flip attackSurface to no, build again — werf sbom get app still reports yes.

What

• After a change to build.sbom.gost, the next build regenerates the SBOM artifact instead of reusing the attached one, including for images with no base or imported SBOM document, where the setting previously reached the checksum through nothing at all.
• The SBOM artifact checksum layout changed, so the first build after upgrading regenerates every image's SBOM once; later builds reuse the cache as before.
• Single-platform SBOM checksums change as well: the target platform now occupies a fixed part slot instead of being appended only when non-empty.
• Two distinct settings can no longer collide through the part separator: a signer identity containing - used to absorb the platform slot boundary.
• Deliberately unchanged: the packages directive (os-pm) and a scratch base stay out of the checksum — both already move the parent stage digest, which is the other half of the SBOM cache key.
• Deliberately unchanged: MergeOpts.Checksum() still covers only base and imported document content.

Why

The SBOM section feeds no stage digest, and the GOST configuration reached the artifact checksum only as a side effect: prepareGostComponents upserts GOST properties into base and imported documents before the checksum is computed. An image with neither left the setting invisible to both halves of the cache key, while gost.Upsert still applied it to the result document at generation time — a silently stale compliance artifact. wiki/pages/sbom-cache-invalidation.md already documented the opposite behavior.

Putting the configuration inside MergeOpts.Checksum() was rejected: GOST is post-processing configuration rather than a merge input, and that channel would count the same setting twice for images that do have base or imported documents.

[reyreavman]

Verification

• Mutation: removed the two GOST parts → gost attack surface, gost security function and GOST config changes checksum even without base and import BOMs failed. Restored the separator-joined encoding → does not collide when a part value absorbs a slot boundary failed; the first version of that case passed under the old encoding too, which is why it was rewritten around a value that absorbs a slot boundary.
• Not run: task test:e2e paths="./test/e2e/sbom/..." labelFilter="sbom" and task test:integration — the host is macOS without WERF_TEST_K8S_DOCKER_REGISTRY; test/e2e/sbom/gost_cache_invalidation_test.go has so far only been compiled.

Review focus

• The checksum layout change invalidates every attached SBOM once. Confirm a one-time regeneration wave is acceptable for the release this lands in.
• pkg/build/sbom_step_checksum_test.go is largely rewritten: two pre-existing cases asserted the old behavior — GOST config does not change unsigned checksum, and one whose name claimed os-pm coverage it never exercised.

Follow-up

• BLOCKER: e2e (labelFilter="sbom") and integration green in CI — the GOST toggle scenario has never executed anywhere.