fix: DatasetCase.metadata() always empty when fetching dataset rows from Braintrust

[James Kennedy (Planview-JamesK)]

Fixes #125

Problem

When running a remote eval that fetches dataset rows from Braintrust (e.g. a dataset assigned in the Playground UI), DatasetCase.metadata() returns an empty map for every row — even though the rows have metadata set. The same experiment works correctly with the TypeScript CLI.

Root Cause

InsertProjectLogsEventMetadata extends HashMap<String, Object> and also declares a separate additionalProperties field. Jackson deserializes unknown properties (all custom metadata keys) into the HashMap base, not into the additionalProperties field. DatasetBrainstoreImpl was calling metadataObj.getAdditionalProperties(), which is always empty.

Fix

Copy entries from the HashMap base via new HashMap<>(metadataObj) instead of calling getAdditionalProperties().

// Before:
Map<String, Object> metadata =
        metadataObj != null ? metadataObj.getAdditionalProperties() : Map.of();
if (metadata == null) metadata = Map.of();

// After:
Map<String, Object> metadata = metadataObj != null ? new HashMap<>(metadataObj) : Map.of();

Verification

Confirmed by running a playground experiment against a Braintrust dataset with metadata fields. After the fix, DatasetCase.metadata() correctly returns the row's metadata keys. Before the fix, metadata_keys=[] for every dataset-fetched row.

[Andrew Kent (realark)]

thanks for flagging this James Kennedy (@Planview-JamesK) , I'll get this merged and released tomorrow