0051. adapters/_common/ Holds Port Semantics No Adapter Should Re-Derive

check_expected — the function that decides whether a stream's current version satisfies the caller's ExpectedVersion — existed verbatim in three store adapters and one testing double. Nothing would have failed if one copy had drifted.

Status

Accepted.

ADR Impact

ADR	Status
0025	Stands. The store ports and the per-technology adapter split it decided are untouched; this only says where logic shared across those adapters lives.
0031	Stands, and supplies the pattern. adapters/_bus/ is the transport-specific case of the same idea; _common/ is the case where the shared code is specific to nothing.

Context

A store adapter's append must answer one question before it writes: does the stream's version satisfy what the caller asserted? The answer depends on ExpectedVersion, OptimisticLockError, and the convention that a non-existent stream is version 0. It depends on nothing about rows, files, or dicts.

Each adapter nonetheless carried its own copy — InMemoryEventStore, SQLiteEventStore, PostgreSQLEventStore, and PartitionedMemoryStore, the last with a docstring explaining that the copy was deliberate, because testing/ must not depend on a specific adapter. That reasoning was sound and the conclusion still wrong: the alternative to depending on one adapter is not copying, it is depending on something that belongs to none of them.

The conformance suite asserts the four adapters agree. It does not stop them from deriving that agreement four times, and the read_category batch-timestamp tie-break had already shown what independent derivation of a conformance-asserted rule produces. This is recurring defect shape #1 caught one step before it becomes a defect: not two implementations that disagree, but four that agree only as long as nobody edits one.

Decision

Add adapters/_common/, a third adapters-internal package alongside _sql/ (dialect-specific) and _bus/ (transport-specific). What lands here is shared by adapters with no technology in common, because it is port semantics rather than storage mechanics. check_expected and describe_expected move there; all four call sites import them.

The standing rule this records: behavior a conformance suite asserts is implemented once. A conformance suite proves the adapters agree; it is not a license for each to derive the agreement separately. When logic depends only on ports and domain types, it belongs in _common/.

_common/ is adapters-internal, not public API — the leading underscore says so, and users get these behaviors through the adapters. It sits in the adapters ring rather than ports/, which contains no implementation code, ever.

testing/partitioned_memory.py imports it. That does not violate the rule its docstring invoked: testing/ must not depend on a specific adapter, and _common/ is by construction the code that belongs to no adapter.

Consequences

One definition to change when the semantics change, and one place a reviewer has to look to know what they are. A new store adapter gets the behavior by importing it rather than by copying the nearest sibling, which is how three of the four copies came to exist.

describe_expected is used by only the two SQL adapters, whose unique-constraint violation surfaces after the insert rather than at the pre-check. It moves anyway: it is the same kind of fact, and leaving it duplicated across two adapters while consolidating its neighbor would be arbitrary.

The package is a place to put things, which is a mild invitation to put the wrong things there. The boundary is dependency-shaped and checkable: if it imports a driver or a dialect, it belongs in _sql/, _bus/, or the adapter itself.