[torbenbrodt]

Check 1 of #564: wire fixtures in adcp/v3's tests are validated against their schemas in the pinned bundle, and the helper fails closed. Test code and AGENTS.md only; no exported Go changes.

What changes

Commit 1: the helper (schema_fixture_test.go).

• The regex list is explicit, and fails closed. test(adcp/v3): validate wire fixtures against their JSON schema #587 already translates the bundle's 8 patterns that Go's RE2 can't run into exact RE2 equivalents (re2Translations, 23 locations in 11 files, listed by schema pointer), so nothing is skipped. TestTranslateSchemaFailsClosed now proves the load fails for an unlisted pattern, an entry for another location or file, a changed pattern text and a bad patternProperties key.
• A missing bundle fails under CI. With CI set, assertSchemaValid and requireSchemaBundle fail instead of logging; locally they still log or skip. Only ci.yml's test job runs these tests, after its bundle download step.
• Known-invalid fixtures. knownInvalidFixtures (empty in this commit; commit 2 adds its entries) records, per fixture, the owning issue, the exact error it must still produce, and a Patch that repairs that one defect. A listed fixture must fail with exactly that error and validate fully once patched, so an entry never hides a second defect, and it fails once the fixture becomes valid. The list records its bundle version (3.2.2) and fails when schemas/VERSION moves.

Commit 2: the call sites. 116 call sites in 20 files (108 assertSchemaValid, 8 through the new assertEncodingSchemaValid; the helper's own tests not counted), up from 57 in 14. Many fixtures turned out to be schema-invalid and are corrected; each expected string changed with its value. The main ones:

Old → new	File
days mon, tue → monday, tuesday; reach unit user → individuals; disclosure overlay_top_left → overlay, persistent → continuous; metric roas → conversion_rate; feedback source buyer → buyer_attribution	generated_core_refs_test.go
idempotency keys under 16 characters lengthened; missing required fields added (buying_mode, proposal_id, total_budget, status, cache_scope, check_id, expires_at, revision, …)	generated_core_refs_test.go, inputs_test.go, responses_test.go
rating system tvpg → tv_parental; landing page required → any; special category sports → championship; readiness ready → good; forecast historical → modeled; selection signal_ids → by_id; publisher_domain, SignalID.Source and image asset fields added	generated_product_refs_test.go
catalog type products → product; preview manifest format_kind; Deployment gains agent_url	inputs_test.go, generated_capabilities_test.go
CreativeInput three-state localization fixtures extended to a full item	creative_input_test.go

Ten fixtures stay invalid because the SDK produces them, each owned by an issue:

• #611 (3): GetProductsResponse.Products is omitempty, so an empty products list is not sent.
• #612 (7): response builders emit no or an empty envelope status (create_media_buy success, get_media_buys, five comply_test_controller results).

Not in this PR

• format is not asserted by jsonschema-go, so a valid fixture can still carry a malformed date-time.
• Fixtures left unvalidated: deliberately invalid inputs (canceled: false against const: true) and a nil optional pointer shown to be omitted say why in a comment; zero-value and error-path tests and stub payloads, for example in idempotent_test.go, optional_test.go, example_optional_test.go, addtool_test.go and validation_test.go (not touched here), are not validated and carry no comment.
• A known-invalid entry whose test is deleted is not detected; it is clutter, not a hidden failure.

Checks run locally

go vet and go test -race ./... in adcp/v3 on Go 1.26 and 1.27 with CI=true; golangci-lint; regeneration reproduces types_gen.go; lint.py --strict; --coverage-max-unreviewed-any 0; the Python generator tests. Each new guard was broken in a copy to show its test fails, including a corrected value put back inside a known-invalid fixture.

Related: #564.

🤖 Generated with Claude Code

[aao-secretariat]

Ladon verdict: Approve

Approve — test-only PR with no blocking or medium findings.

This PR touches 11 files (all *_test.go plus AGENTS.md); no production, schema, or generated-type changes. It hardens the adcp/v3 schema-fixture helper and extends validation from 57 to 116 call sites.

Checked and clean:

• schema_fixture_test.go known-invalid control flow fails closed; one entry can't mask a second defect; assertKnownInvalid/translateSchema fail-closed on unlisted RE2 patterns.
• Global-map mutation in TestAssertSchemaValidReports is safe — no t.Parallel() in the package, subtests awaited serially, no race.
• Fixture value corrections are hand-built test inputs with want strings updated in lockstep; no assertion weakened.
• Ten remaining-invalid fixtures are pre-existing SDK wire bugs tracked to issues (#611, #612), not introduced here.
• No touch to types_gen.go, tmproto signing, identity-agent TEE, protocol-managed skills, or any wire-shape/exported-API — changeset/minor-pathway rules don't apply.

Gate check: gated_paths is false, no high-risk flag, no no-auto-approve team match, no findings. Rows 1–8 do not fire; falls through to row 9 (approve). Note review_decision is REVIEW_REQUIRED, but row 2 only applies when gated_paths is true, which it is not.