Release data-designer-retrieval-sdg 0.2.0

[shan-nvidia]

Summary

• Bump data-designer-retrieval-sdg from 0.1.0 to 0.2.0.
• Include the dependency-security updates merged in Fix vulnerable plugin dependency resolutions #56: NLTK >=3.10.2,<3.11, PyArrow >=24,<25, MCP >=1.29,<2, and Starlette >=1.4.1,<2.
• Require data-designer>=0.9.1 in package and catalog metadata.
• Rely on Data Designer 0.9.1's upstream Parquet schema-unification fix instead of publishing a retrieval-specific score-normalizer processor.
• Refresh the workspace lock, plugin docs, and catalog metadata for the two retrieval SDG extensions.

The existing v0.1.0 release predates native Data Designer resume (#51), the stable recipe APIs (#54), reproducible run configuration (#55), and the dependency upgrades in #56 and #58. This release gives that combined baseline a unique installable identity so downstream consumers such as the Nemotron embedder recipe can pin it reproducibly.

Mixed-score Parquet fix

The quality schema declares qa_evaluations.evaluations[].overall.score as a float, but model output can encode an integral score as 9 and a fractional score as 9.5. Separate checkpoints can therefore infer int64 and double schemas, causing final profiling/export to fail:

pyarrow.lib.ArrowInvalid: Float value 9.500000 was truncated converting to int64

This was fixed upstream in Data Designer #858 and released in Data Designer 0.9.1. The reader now unifies compatible Parquet schemas before loading the dataset. This PR therefore removes the temporary qa-evaluation-normalizer implementation, registration, tests, and documentation, and raises the retrieval package's minimum Data Designer version to 0.9.1. No new public plugin is introduced for this storage-layer issue.

Validation

• 309 tests across devtools and all plugins in the repository's package-scoped layout
• Retrieval SDG suite under Data Designer 0.9.1: 136 passed
• Exact mixed 9 / 9.5 checkpoint reproduction reads as [9.0, 9.5] through Data Designer 0.9.1 without a plugin processor
• Ruff lint and formatting checks
• All installed Data Designer entry points validate; retrieval SDG exposes only document-chunker and embedding-dedup
• Strict documentation build and generated-doc check
• Catalog validation and SPDX header validation
• ddp check-release data-designer-retrieval-sdg 0.2.0
• Build the 0.2.0 wheel and source distribution from candidate commit b9f17dd
• Current Twine validation on both artifacts
• Verify isolated retrieval-package resolution selects Data Designer 0.9.1
• Inspect the built wheel and verify the CLI plus the two intended retrieval entry points
• Prior credentialed generation, interruption/resume, conversion, Nemotron Stage 0 local, and Nemotron Stage 0 Docker qualification completed on the pre-release 0.2.0 candidate

Candidate commit: b9f17dd91c937a328f9a63a5cd4ca6d86f6e48e1

Candidate wheel SHA256: 1c448d318414e3a2a37fc61db3e07ace1ff7c838f9769355eb92faaca103ff06

Candidate source distribution SHA256: b702bbf194e1fd625bbf01423e9239c49c8224dcce7ebf31d80b9fc9bc9a4224

The earlier launcher qualifications exercised the 0.2.0 package API and pipeline with Data Designer 0.9.0. The final dependency-only delta was validated against 0.9.1 with the complete package suite and the exact mixed-schema regression that required the upgrade.

Docker qualification used image digest sha256:1dc787f5c6264fcc1c99809f99b84823e73ed4588d5a581b94290fc2a8fecff8. It produced one JSONL record with plugin provenance 0.2.0; the exact API key was absent from output and saved job configs, the validation container exited, and the original remote Nemotron checkout remained clean.

Pre-Publish Gates

• Run a small credentialed generation job using the 0.2.0 candidate.
• Interrupt and resume that job using Data Designer native resume.
• Verify the top-level Nemotron Stage 0 local launcher.
• Verify the top-level launcher in the exact Nemotron Stage 0 container.
• Verify Data Designer 0.9.1 against the exact mixed-schema failure.

All pre-publish validation gates are complete. After merge, rebuild and hash the artifacts from the merged commit before creating the data-designer-retrieval-sdg/v0.2.0 release. Repeating local and Docker checks with that immutable release plus qualifying a non-Docker remote profile remain part of the subsequent Nemotron migration.

Checklist

• The package exposes two valid retrieval SDG extensions.
• assert_valid_plugin(plugin) passes for every installed entry point.
• Unit tests pass in package-scoped environments.
• The built wheel contains the CLI and both retrieval plugin entry points.
• Plugin docs are current.
• Documentation builds strictly.
• catalog/plugins.json advertises the Data Designer >=0.9.1 compatibility floor.
• Existing per-plugin CODEOWNERS remains in place.
• NVIDIA SPDX headers pass.