The reader must receive 7 after the required conflicts are resolved. The peer's CD reaches the interconnect; the requesting reader receives data through R. Memory's older 0 is not made current by being faster to return.

No. Retain the grant's completer/type context and wait for both prerequisites. The credited retry uses AllowRetry=0 and the required PCrdType. Admission still does not mean read completion or a fixed latency guarantee.

Independent fields occupy one coherence granule, so writers can invalidate or transfer permission even though the fields are logically separate. Confirm line-addressed traffic and access ownership. Separation/alignment, partitioning or batching may help, with footprint trade-offs; false sharing alone does not explain incorrect data in an otherwise correct system.

Common trap: Fixing with a faster NoC only or treating extra coherence traffic as proof of corruption.

In the abstract plain-MESI model, the line is valid and clean and other caches may also have it. Reads are allowed; a local write must first obtain the required unique permission and resolve other copies. Requesting an upgrade is not already completing it. ACE's response/state vocabulary needs its own mapping.

Common trap: Confusing Shared with Exclusive or writing before permission arrives.

The message classes carry different roles and dependencies, allowing implementations to provision and route them separately. Deadlock freedom still requires the specification's dependency and progress rules plus sufficient implementation resources; separate channels alone are not proof.

Common trap: Treating CHI as 'AXI with snoops' or assuming channel separation automatically prevents deadlock.