Explanation: rejected alternatives¶

Context-based repository lookup¶

Rejected because context becomes a service locator, dependencies become invisible, missing wiring fails at runtime, and transaction ownership becomes ambiguous.

Replacement: models retain a private session attachment.

Public repository interfaces¶

Rejected because they undermine the intended model.Save(ctx) Active Record API and duplicate sqlc abstractions.

Replacement: public collections retrieve models; private stores adapt sqlc.

Public restore or mark-persisted methods¶

Rejected because callers could forge lifecycle state and bypass invariants.

Replacement: hydration and snapshot synchronization remain private to the bounded-context package.

Package per relation traversal¶

Rejected because of import cycles, unbounded traversal paths, duplicated concepts, ambiguous semantics, and inability to hydrate private target state.

Replacement: public relations share the model package; internal loaders may use nested directories.

Automatic cascading persistence¶

Rejected for the first release because it requires hidden query ordering, implicit transactions, cycle detection, cascade policies, and difficult error recovery.

Replacement: reject unsaved related models without keys and require explicit saves.

Dynamic SQL generation¶

Rejected because it would create a second query engine and weaken sqlc's static-analysis boundary.

Replacement: typed scopes map to predefined queries and parameters.

Mandatory identity map¶

Deferred because it introduces memory retention, session lifetime, stale-object policy, transaction isolation, and concurrency concerns.

Replacement: document independent mutable snapshots and hydrate known inverse instances where possible.