cursus.registry.step_names_base

Base step-name registry — the single source of truth for step naming across config, builders, and specifications.

This is a deliberately dependency-free leaf module: it builds the flat step table STEP_NAMES from the per-step .step.yaml registry: blocks (via build_registry_from_interfaces) and derives the CONFIG_STEP_REGISTRY / BUILDER_STEP_NAMES / SPEC_STEP_TYPES mappings. The workspace-aware access layer (step_names.py) and the hybrid UnifiedRegistryManager (hybrid/manager.py) both read the raw data from here — keeping the data in this leaf (rather than in step_names.py) is what breaks the otherwise-circular import between the access layer and the manager.

Source of truth (FZ 31e1/31e1f Final Phase, 2026-06-28): the standalone registry/step_names.yaml table was DELETED. The registry is now derived SOLELY from the .step.yaml registry: blocks + a 3-row _EXTRAS map (in interface_registry_loader) for the interface-less abstract steps (Base / Processing / HyperparameterPrep). To add or edit a step, edit its .step.yaml registry: block — there is no separate table. A golden snapshot (tests/registry/step_names_registry_snapshot.json) gates drift.

(Formerly step_names_original.py; renamed because it is the live source, not a backup.)

merge_pack_registry(pack_rows)

Merge an external step-pack’s registry rows ON TOP of the package registry, in place.

ADD-ONLY overlay: package steps are the floor and are never removed. STEP_NAMES is mutated in place (so import-time references stay live) with pack_rows layered on top, then the derived globals are rebuilt. This is the LOW-LEVEL merge — the access layer (cursus.registry.step_names.refresh_registry()) calls this AND then re-syncs the hybrid manager so the catalog sees the plugin steps.

Parameters:

pack_rows (Dict[str, Dict[str, str]]) – {canonical_name: {config_class, builder_step_name, spec_type, sagemaker_step_type, description}} derived from the pack’s .step.yaml files.

Returns:

{name: "collision"} for any pack name that shadowed an EXISTING package step — the caller decides whether to warn. Empty when every pack row is genuinely new.

Return type:

Dict[str, str]