Layer dependencies

Canonical rules: Principles → Framework layers.
Ownership map, anti-leak rules, and “where to put new logic” live there.

Canon

drtml → contracts / storage.plan → db | storage → processing → telemetry | integration

Reverse edges (storage → processing|telemetry, telemetry → processing, heavy imports from contracts/) are forbidden.

Connection modules read environment through runtime/env.py; db does not depend on integration.

contracts/ (examples)

Module	DTO
`saturation_slices.py`	saturation / homogeneity slices
`dataset_metric_rows.py`	rows for metric evaluate
`worker_metric_delta.py`	multi-worker metric merge
`mapping_build.py`	mapping build result and stats
`reduce_merge.py`	reduce/merge result and stats
`embed_train_stats.py`	embed-train runtime statistics + public dict
`embed_train_metric_catalog.py`	live/final embed-train gauge names
`observation_window.py`	bounded observation-window DTO
`dataset_metric_window.py`	dataset metric checkpoint window

Registries

Registry	Where
Dataset write/read	`storage/dataset/backend_registry.py`
Pattern / mode runners	`runtime/pattern_runners.py`
Parallel transports	`partition_loop/parallel/backend_registry.py`
Metric methods	`algorithms/dataset_metrics/registry.py`

CI

# from DrToller.Framework repo root
pytest dev/tests/architecture/test_layer_import_boundaries.py -q
pytest dev/tests/storage/test_dataset_backend_adapter.py -q