Expose Datagen history reads and retain folded codec metadata

[YangjunZ]

Summary

• Expose Python item_events / root_events through embedded and remote stores, preserving overwritten values, per-item ordering and lazy blobs. Encode fan-out IDs safely in HTTP paths.
• Retain field_type / codec_version in folded fields. SET metadata follows the winning value; APPEND rejects mixed codecs instead of mislabeling accumulated values.
• Refresh cached server handles before raw-history scans so external WAL merges cannot leave a stale, nonempty prefix. Add regression coverage and document read consistency.

Compatibility

No event-schema migration or changes to Context, Rollout, Generic, or shared MemWAL behavior. Python field dictionaries gain metadata keys; older server responses expose missing metadata as None, not an inferred codec.

Rust users must update DatagenFieldState variants, DTO literals and custom DatagenStoreApi implementations. Mixed-codec append histories now fail to fold. Deploy the updated server before using the new item-history endpoint.

Validation

• Full Rust workspace: 375 passed, 25 ignored. Final server regressions passed after reproducing and fixing the stale-prefix case.
• Full Python suite, including real HTTP and local/S3 persistence: 229 passed, 1 failed, 1 skipped, 1 xfailed.
• Formatting, Ruff, workspace Clippy and default-feature Clippy passed.
• Native Python module and server built locally; independent Opus 5 review and focused follow-up completed.

Existing baseline issues, both reproduced on pristine 8fe27b2 and left out of scope:

• test_persistence.py::test_image_round_trip reads zero base-table rows after the image write.
• With optional Pillow installed, Pyright reports the existing Image.get_format_mimetype access at api.py:210.

The Python skip is opt-in GCS integration; the expected failure is the existing MemWAL/base-version time-travel test. No pre-commit hook or configuration is installed in this repository.

🤖 Generated with MAI Agents