Review: schema-aware context publish

Well-scoped change. The generated context.pb.go looks like a clean protoc regeneration (consistent msgTypes / goTypes / depIdxs, NumMessages 12 → 14, exporters shifted correctly), Publish keeps its raw-bytes behavior, and the new unit tests cover both the raw and schema-aware paths with a capture stub. A couple of inline points on public-API consistency and runtime compatibility — neither is a hard blocker.

Non-blocking notes:

• .ci/helm.sh ci::verify_total_processed: good fix changing the match from *"$total" to *" ${total}" — the old suffix match would false-positive (e.g. total=1 matching a metric value of 121).
• Example decodeStudent relies on trial-and-error decoding across two Avro schemas. Binary Avro isn't self-describing, so decoding with the wrong schema can succeed with garbage instead of erroring. It's fine here since CI asserts exact output values (grade+1), just flagging the latent fragility for anyone copying the example.

Re-review of "Refine context publish schema API" (fca7030)

Both points from the previous pass are resolved — thanks for the quick turnaround:

• API consistency ✅ — PublishWithSchema now takes a hand-written value struct PublishMessageSchema (with SchemaData back to string) and maps to the wire proto internally via publishMessageSchemaToProto, mirroring the SinkSchema pattern and keeping the generated type off the public surface. Example and unit test are updated accordingly, and there are no lingering references to the old pointer signature.
• Runtime compatibility ✅ — the PublishWithSchema godoc now documents that a runner consuming PulsarMessage.schema is required (referencing generic-runtime#42) and that older runners silently ignore the proto3 field.

One optional minor note inline. LGTM otherwise.