1 datasets.bundles

datasets.bundles

Cross-store dataset composition; import each workflow from its owning leaf.

1.1 Package guide

A bundle selects one immutable VIN root and explicit rollout stores. The dataset guide explains the dependency direction, executable usage and verification. inspection hides lightweight acquisition and compatibility decisions; readiness hides Q_H admission; deep_scan hides the opt-in array scan and inventory acquisition. export consumes those completed products without acquiring them again.

_payloads is private to this package and implements shared immutable JSON projections. Empty package markers deliberately expose no symbol catalog. Scientific populations, hard masks, generation hashes and read-only behavior remain in the workflow docstrings and tests; plotting belongs to callers.

corpus composes explicit rollout membership with multiple root candidates. DatasetCorpusSelection records root candidates and optional exact manifest/path binding overrides. Missing or ambiguous dependencies remain visible members; only aliases of the same physical root deduplicate automatically. Its V2 generation covers every selected alias, root, rollout and override. Inspection, QH admission and an actual batch preview remain separate dispatched operations.

For training, QhDatasetConfig.actor selects one root and actors=(...) selects multiple roots; these inputs are exclusive and normalize to the same qualified join and provenance path. Root-local sample indices are qualified by manifest hash in joins and CPU chain keys. Existing map-style DataLoader shuffle samples rollout chains uniformly; worker-local readers reopen their own handles. Physical storage shards do not need permanent worker assignments.

One actor-state contract describes input representation compatibility. Exact TRAIN identity hashes that contract together with complete TRAIN provenance; receipt-bound stage provenance seals roots, source rows and rollout manifests. Voxel extents, EVL shapes, storage capabilities/versions, point features, profiles, scene disjointness and supervision admission remain checked. Current bundles require the original training receipt, including exact stage populations.

Current loading rejects retired contract and receipt formats. Original artifacts and historical reports remain unchanged. The supported current path is a separately admitted fit with its own receipt and checkpoint. Synthetic composition tests cannot establish trained-checkpoint admission for a new multi-root population. Legacy supervision remains rejected by the default dense-valid readiness objective unless its original persisted support meets that objective.

# datasets.bundles { #aria_nbv.datasets.bundles } `datasets.bundles` Cross-store dataset composition; import each workflow from its owning leaf.  ## Package guide A bundle selects one immutable VIN root and explicit rollout stores. The [dataset guide](datasets.qmd) explains the dependency direction, executable usage and verification. `inspection` hides lightweight acquisition and compatibility decisions; `readiness` hides Q_H admission; `deep_scan` hides the opt-in array scan and inventory acquisition. `export` consumes those completed products without acquiring them again. `_payloads` is private to this package and implements shared immutable JSON projections. Empty package markers deliberately expose no symbol catalog. Scientific populations, hard masks, generation hashes and read-only behavior remain in the workflow docstrings and tests; plotting belongs to callers. `corpus` composes explicit rollout membership with multiple root candidates. `DatasetCorpusSelection` records root candidates and optional exact manifest/path binding overrides. Missing or ambiguous dependencies remain visible members; only aliases of the same physical root deduplicate automatically. Its V2 generation covers every selected alias, root, rollout and override. Inspection, QH admission and an actual batch preview remain separate dispatched operations. For training, `QhDatasetConfig.actor` selects one root and `actors=(...)` selects multiple roots; these inputs are exclusive and normalize to the same qualified join and provenance path. Root-local sample indices are qualified by manifest hash in joins and CPU chain keys. Existing map-style DataLoader shuffle samples rollout chains uniformly; worker-local readers reopen their own handles. Physical storage shards do not need permanent worker assignments. One actor-state contract describes input representation compatibility. Exact TRAIN identity hashes that contract together with complete TRAIN provenance; receipt-bound stage provenance seals roots, source rows and rollout manifests. Voxel extents, EVL shapes, storage capabilities/versions, point features, profiles, scene disjointness and supervision admission remain checked. Current bundles require the original training receipt, including exact stage populations. Current loading rejects retired contract and receipt formats. Original artifacts and historical reports remain unchanged. The supported current path is a separately admitted fit with its own receipt and checkpoint. Synthetic composition tests cannot establish trained-checkpoint admission for a new multi-root population. Legacy supervision remains rejected by the default dense-valid readiness objective unless its original persisted support meets that objective.