test(java): jsonb open-bag runtime-write gate + api-contract harness (#98)

[dmealing]

Completes the Java side of the jsonb-open-bag (field.string @dbColumnType:jsonb) cross-port consistency work. Two mandates.

Mandate 1 — OMDB runtime write path: did it have the gap? No (well-evidenced).

The TS gap (PR #110 found) was a runtime write validator rejecting a posted object. OMDB has no analogous gap:

• OMDB's open-bag write path correctly serializes + binds a JSON-text value: isOpenJsonbField → serializeOpenJsonb → bindJsonbParameter (Postgres setObject(.., Types.OTHER)).
• OMDB is pure data-access (ADR-0015) — there is no write-side validator to reject anything.
• The persistence corpus never exercised the open bag through the runtime write codec: asset-uuid-roundtrip.yaml seeds payload via raw SQL and only READS; roundtrip-all-types.yaml writes the typed @storage:jsonb path (Settings), not the open @dbColumnType:jsonb bag. OpenJsonbWriteRoundtripTest fills that gate — it INSERTs an Asset whose payload is JSON text through ObjectManagerDB.createObject (NOT raw SQL) and reads it back parsed.

The "object is rejected" symptom is not an OMDB write-path defect. OMDB's ValueObject model is statically typed by field subtype (ADR-0017), so a structured value assigned to a field.string is coerced to its java toString() at set-time (DataObjectBase._setObjectAttribute → DataConverter.toType(STRING, …)), identical for every string field — by design. Verified empirically: a Map bound to the bag binds the malformed text {a=1, b=2}, which Postgres rejects with invalid input syntax for type json. The cross-port "write an OBJECT" half lives at the DTO/serialization boundary (the generated DTO is Object, #103): a consumer serializes the posted object to JSON text before it reaches OMDB's field.string, and the read side re-parses it.

No OMDB production change was needed — hence the test() prefix. (GenericSQLDriver is byte-pristine; I prototyped a writeValueFor raw-read fix, proved via empirical RED that set-time coercion defeats it for ValueObject, and reverted it.)

Mandate 2 — Java api-contract jsonb harness (both lanes)

Mirrors the single-entity Author + m2m/tph harnesses for the new fixtures/api-contract-conformance/jsonb/ corpus:

• Reference lane — JsonbReferenceServer (JDK HttpServer + Postgres testcontainer, documents table with payload JSONB).
• Generated lane — GeneratedJsonbControllerHarness (codegen-spring DocumentController/DocumentDto compiled + booted over MockMvc behind the in-memory DocumentRepository seam).

The generated DocumentDto types the bag as Object (#103), so a posted JSON object binds and reads back parsed, never a JSON-encoded string. This LOCKS the API-boundary contract.

Test evidence (Testcontainers Postgres)

OpenJsonbWriteRoundtripTest                  1/1 green
JsonbApiContractConformanceTest (reference)  1/1 green
JsonbGeneratedApiContractConformanceTest     1/1 green
ApiContract*/M2m*/Tph* (existing)            48/48 green (no regression)

Deferred

• Kotlin + C# api-contract jsonb harnesses (sibling worktrees).
• A cross-port persistence-conformance open-bag op: roundtrip scenario (would require all five ports' roundtrip writers — out of this Java-scoped change).

Part of #98

🤖 Generated with Claude Code