ProtoMolt court-to-search contract demonstration

This is an executed sample, adapted from ProtoMolt's MessageProjectionTest.
The filing record is synthetic test data: 24-cv-00117, State v. Example.
The transformations, rejection, two loopback gRPC calls, chunking, and model
inference were actually executed. result.json contains the captured output.

The DocumentSink RPC validates and acknowledges a document. It does not write
an index. ProtoOptionsIndexingHintSource reads the real index annotations;
PolicyDerivation executes the declared chunking and embedding policy. The
sample yields one title chunk and a 384-component float vector. COSINE is the
declared search similarity; this demonstration does not perform a search.

Documentation is carried as protobuf comments and retained in the descriptor
set with protoc --include_source_info. This is not a claim that arbitrary
registered RPCs acquire their own HTTP/OpenAPI endpoints.

Files:
  court-search.proto  Complete compiled source and destination service contract
  ContractDemo.java   Executed Java program, including the two local RPC servers
  run.sh              Compile and execute against a supplied ProtoMolt classpath
  result.json         Complete captured output including all 384 coordinates
  provenance.json     Source revision, model fingerprint, runtime artifact hashes

Reproduce with Java 25, protoc and the ProtoMolt source/runtime artifacts at:
  dc49800cd66950356a883d99141e00427df6ad65
Supply the runtime classpath containing projection, protobuf validation,
search-index-spi, search-embedding, gRPC Netty/protobuf/stub, protobuf-java-util,
Gson, and their dependencies. The recorded artifact filenames and SHA256 hashes
are listed in provenance.json. run.sh compiles the chunking module directly
from that checkout into a temporary directory; it does not edit the checkout.

Use an already-running TEI gRPC server serving:
  sentence-transformers/all-MiniLM-L6-v2
The recorded run used TEI 1.9.3, float16. Model weights and cache revision are
fingerprinted in provenance.json. Hardware/runtime differences can cause
small numerical differences. The Java program checks the server model ID,
actual returned vector length, finite coordinates, L2 norm, expected mapped
values, successful validation, and INVALID_ARGUMENT for title "X".

Set PROTOMOLT_SOURCE to the pinned checkout, PROTOMOLT_CLASSPATH to its runtime
classpath, and TEI_TARGET to your existing TEI gRPC host:port. Then run:
  bash run.sh /tmp/court-search-result.json

The script retains temporary descriptor sets and classes and prints their
location. All service listeners are loopback-only on ephemeral ports and are
shut down at completion. No credentials or production endpoint are bundled.