Named-result carriage

Generated named queries have one semantic result shape fixed by the compiled query plan. RiffDB may carry that shape through a negotiated wire encoding, but application code and agents cannot select the physical encoding or provide field ordinals, layouts, or schemas.

The closed encoding order is:

1. LEGACY_RECORDS
2. COMPACT_V1
3. PACKED_V1

Each request advertises a contiguous prefix. Servers return exactly one arm, and clients reject an unadvertised, mixed, malformed, over-bound, or identity-mismatched arm before releasing a partial result.

Current generated clients advertise through COMPACT_V1 for eligible compiler-covered queries. PACKED_V1 exists in the additive protocol and strict driver compatibility codec, but generated clients neither select it nor carry dormant direct decoders because its candidate implementation missed both the performance and generated-size gates. There is no application configuration switch for it. Queries without a complete safe cover continue to use legacy records.

One-level bounded expansion also uses the existing name-addressed legacy value carrier. Each driver record owns one compiler-named list of target records; the executor retains the driver key while assembling the flat authorized target observations, then releases only the nested typed result. Generated Rust, Go, TypeScript, and Python decoders accept exactly that compiled record shape and reject missing, extra, malformed, or over-bound nested values. MCP and CLI render the same named list and cannot request a flat mode, index, driver, fan-out, or per-driver cursor.

The local driver protocol is V3. The V3 host remains read-compatible with V1 legacy and V2 compact clients. Remote trust, authorization, freshness, cancellation, retry, and uncertainty semantics do not vary with carriage.