refactor: make Record source-agnostic with RecordSource discriminated union

[rorybyrne]

Summary

Make the Record aggregate source-agnostic so it can be created from any pathway (deposition, harvest, fork, import) without coupling to DepositionSRN.

Motivation

The harvest domain needs to create records directly from bulk-ingested data without going through the deposition lifecycle. Currently, Record has deposition_srn: DepositionSRN as a required field, and the entire downstream chain (RecordPublished event, InsertRecordFeatures handler, FeatureService) is coupled to depositions.

Changes

1. RecordSource discriminated union (new shared model)

class RecordSource(ValueObject):
    type: str   # "deposition" | "harvest" | "fork" | "import"
    id: str     # idempotency key

class DepositionSource(RecordSource):
    type: Literal["deposition"] = "deposition"

class HarvestSource(RecordSource):
    type: Literal["harvest"] = "harvest"
    harvest_run_srn: str
    upstream_source: str  # "pdb", "geo"

class ForkSource(RecordSource):
    type: Literal["fork"] = "fork"
    source_node: str

class ImportSource(RecordSource):
    type: Literal["import"] = "import"
    source_node: str

2. Record aggregate (simplified)

class Record(Aggregate):
    srn: RecordSRN
    convention_srn: ConventionSRN
    source: RecordSource  # JSONB in DB
    metadata: dict[str, Any]
    published_at: datetime

Remove: deposition_srn, indexes.

3. Records table schema change

Alembic migration:

• Drop deposition_srn column + its index
• Drop indexes column
• Add convention_srn TEXT NOT NULL column
• Add source JSONB NOT NULL column
• Add functional unique index: UNIQUE(convention_srn, source->>'type', source->>'id')

No backfill or backwards compatibility needed — no production data exists.

4. RecordPublished event

Replace deposition_srn: DepositionSRN with source: RecordSource. Keep features_dir for downstream feature insertion.

5. RecordService

Update publish_record() to accept source: RecordSource instead of deposition_srn. Add bulk_publish() method for batch record creation.

6. HookInputs (validation port)

Replace deposition_srn: DepositionSRN with generic work_dir: str. Callers set the path.

7. FeatureService / InsertRecordFeatures handler

• Handler uses convention_srn to look up hooks from convention (instead of receiving hooks from event)
• FeatureService reads features from features_dir passed through the event

Files touched

• domain/record/model/aggregate.py — Record aggregate
• domain/record/service/record.py — RecordService
• domain/record/event/record_published.py — RecordPublished event
• domain/shared/model/source.py — new RecordSource models
• domain/validation/port/hook_runner.py — HookInputs
• domain/feature/service/feature.py — FeatureService
• domain/feature/handler/insert_record_features.py — InsertRecordFeatures
• infrastructure/persistence/tables.py — records table
• infrastructure/persistence/record_mapper.py — mapper
• infrastructure/k8s/runner.py — K8sHookRunner
• Alembic migration

Blocked by

Nothing — this is a standalone refactor.

Blocks

#104 (Harvest domain implementation)