Feed and WorkUnit

Parse vs runtime

Layer	Where
Parse StepFeedPlan	drtml/models/feed/
Runtime iteration	storage/feed/

WorkUnit

Unit of work for partition_loop: primary rows (+ attach join when needed).
The processor sees already materialized logical dicts.

Partition semantics (do not break)

• partition.source=file + group_by — groups inside one shard.
• partition.source=rows + group_by — groups across all part files; partition_id without shard.
• corpus_scan — explicit partition in DRTML; framework does not invent domain columns.

Modules

File	Role
shard_plan.py	primary shards + work-unit counts (one pass)
shard_scope.py	filter shards for a worker
executor.py	iteration API only
attach_store.py	join cache; RoutingIndexLookup.load only here
keys.py	join/group/partition keys

Multi-worker

Driver assigns primary shard metadata round-robin; the worker reads feed locally with shard_scope.
Shard counting lives in shard_plan, not processing/parallel/.