DUR041 carries the real state of conformance coverage

[gplanchat]

Text only — one ADR and one docblock. No behaviour, no test changed; the 134 conformance tests pass, unchanged.

What was false

DUR041 describes two arrangements in the present tense although they were never written:

The Temporal adapters answer to a live server and belong to the integration suite; they run this tier there, not in unit.

Temporal's read-through store gets checked for the first time.

Neither TemporalJournalEventStore, nor TemporalReadThroughEventStore, nor TemporalWorkflowRunCatalog extends a conformance suite. No tier runs against them — not the port tier, not the replay tier, not in unit, not in integration.

The docblock of EventStoreReplayConformanceTestCase repeated the same statement ("an adapter backed by a server — both Temporal stores — only extends EventStoreConformanceTestCase"), so a reader who checks the ADR against the code lands on the same claim and concludes it is true.

What was not false

The status said "implemented for all four ports". That is accurate: all four ports do have their suite. What was missing is that adapter coverage is a different question, and that nothing in the ADR gave it.

And a pair only exists if the backend implements the port. Temporal implements two out of four — EventStoreInterface (twice) and WorkflowRunCatalogInterface. WorkflowMetadataStore and ChildWorkflowParentLinkStoreInterface have no Temporal implementation, as the ADR's context table already says with dashes: those are not gaps.

The real state, now in the ADR

Port	Memory	DBAL	Illuminate	Temporal
EventStoreInterface	two tiers	two tiers	two tiers	implemented twice, runs nothing
WorkflowMetadataStore	runs	runs	runs	not implemented
ChildWorkflowParentLinkStoreInterface	runs	runs	runs	not implemented
WorkflowRunCatalogInterface	runs	runs	runs	implemented, runs nothing

Twelve pairs exist and are covered, two exist and are not — the two Temporal ones, the only backend whose storage shape differs from the reference. The three covered backends share a journal and a table: the parity they prove is the one that was never in doubt. The suite is proven where it was free and unproven where it matters, which is the opposite of what the ADR set out to do.

The shape of the correction

The two faulty sentences are kept and annotated, not rewritten. They state the decision, which holds: splitting into two classes is sound, and an adapter backed by a server should indeed extend only the port tier. It is the present tense that did not hold. One observation is worth keeping as is:

A bridge that runs neither half was not anticipated by the split, and that is precisely the oversight it was meant to make visible. The declaration was never made, so there was nothing to notice.

Follow-up

Filling the two pairs is the subject of #269 (backend-data-parity), whose §5 writes them and whose §5.3 adds a CI job that fails on a missing pair — so the grid does not end up incomplete a second time without anyone seeing it.