[Rust] MessageRegistry::from_descriptors for sibling/imported types (PR #461 feedback)

[pixie79]

Summary

Adds MessageRegistry::from_descriptors (behind the zeroparser feature) and addresses the review feedback from upstream databricks/zerobus-sdk#461.

from_descriptor only recurses nested_type, so it registers the root and its nested children but never the sibling/imported top-level messages from other files in a FileDescriptorSet. A root field typed e.g. .google.protobuf.StringValue then misses on registry.get() and decode fails with UnknownTypeName. from_descriptors(root, others) closes that gap by additionally registering each other top-level message under its true package-qualified FQN.

Feedback addressed

1. True e2e coverage — create_registry_for_version now builds the registry via from_descriptors, threading every top-level message in the compiled FileDescriptorSet (including the imported google.protobuf types) under its package-qualified FQN. The former negative test test_google_protobuf_types_not_in_registry is replaced by test_google_protobuf_types_decode_success, which decodes a GoogleProtobufTypesMessage and asserts Timestamp, Duration, and every wrapper type (StringValue, Int32Value, Int64Value, UInt32Value, UInt64Value, FloatValue, DoubleValue, BoolValue, BytesValue) plus repeated Timestamp resolve and decode correctly — for both proto2 and proto3, using the real descriptor-set path.
2. Borrowed strings — API is now others: &[(&str, &DescriptorProto)]; call sites/tests updated.
3. Trimmed rustdoc — the verbose doc comment on from_descriptors is reduced to a concise summary.
4. Benchmarks — ran the existing Criterion parser bench before/after; see below.

Tests

cargo test -p databricks-zerobus-ingest-sdk --features zeroparser --lib zeroparser::registry   # 6 passed
cargo test -p databricks-zerobus-ingest-sdk --features zeroparser --test zeroparser_e2e        # 98 passed
cargo clippy --all --all-features -- -D warnings                                               # clean
cargo fmt --check                                                                              # clean

Benchmark evidence

cargo bench -p databricks-zerobus-ingest-sdk --features zeroparser --bench zeroparser_parser_bench -- --save-baseline before   # on 80fcd0b (base)
cargo bench -p databricks-zerobus-ingest-sdk --features zeroparser --bench zeroparser_parser_bench -- --baseline before        # with this change

The change touches only from_descriptors and the test-only registry helper — none of it is in the benchmark's decode b.iter loop, so no code-attributable change is expected. The comparison confirms this: all three decoders (prost, prost_reflect, and zeroparser) moved together by a uniform ~40–50% between the two runs, the signature of environmental variance (the base run executed under higher load) rather than a real effect on the byte-identical parser hot path. Net: no regression in the parser.

🤖 Generated with Claude Code