[granthamtaylor]

Registers relflow-2026-10-21 (relflow-latest) as a native system with a frozen schema, feature query, and training recipe for each of the 21 entity tasks. It trains on the supplied inner split, selects checkpoints by complete-validation AUROC or MAE, and predicts the masked outer table in its original row order using the supplied outer database. It keeps the selected inner checkpoint without a train-plus-validation refit.

The PR contains the runnable adapter, required task resources, focused tests, and dependency/setup documentation only. RelFlow is pinned to a public source revision and the frozen MiniLM encoder to a Hub revision. Its Arrow requirement uses a separate optional development environment. No experimental controllers, logs, data, checkpoints, or historical result tables are included; baseline_results/ is unchanged.

Recipes were selected using development validation data. This single-SDK integration is a new comparison and makes no claim to reproduce historical numbers or a fixed parameter count. Official test evaluation and fixed-hardware retraining remain for maintainer review.

Validation:

• RelFlow integration: 69 tests passed, including actual tiny CPU classification/regression fits, checkpoint reload parity, masked inference, and optional-backend isolation.
• Existing suite in the default environment: 540 passed, 11 skipped. macOS used OMP_NUM_THREADS=1 and the Torch OpenMP library path; the RT checkpoint test needed safetensors installed in the temporary test environment.
• A separate two-update CPU smoke test passed with the pinned MiniLM encoder, cached text, checkpoint reload, and masked inference.
• All 21 SQL queries bound against the lab's real source schemas and schemas after simulated official empty-column cleaning; footer/schema checks only, no source rows or task/test tables read.
• Both package builds and all pre-commit checks passed.

See RelFlow setup and protocol. No official test scores have been evaluated for this submission.