Define the relation before reading clocks
Happens-before expresses an ordering justified by causality or specified synchronization rules. In Lamport's distributed-system model, local process order, a message send preceding its receipt, and transitivity generate the relation. Not every pair of events is ordered, so the result is a partial order. Events that are unrelated by it are concurrent in the model, even if an observer assigns them different wall-clock timestamps. A numerical timestamp comparison from independent machines does not by itself demonstrate a causal connection or a usable synchronization guarantee.
Following a message chain
Suppose process A computes a result, sends a message containing it, and process B records the result after receiving that message. The computation precedes the send locally, the send precedes the receive, and the receive precedes the record locally. Transitivity therefore orders the computation before the record. An event at a third process C with no connecting communication is not automatically ordered by this chain. Being unordered does not mean that the events occurred at exactly the same physical instant; it means that the defined relation does not establish their order.
Shared-memory rules are language-specific
Shared-memory language models also use happens-before to describe ordering, but the operations that establish the relation are determined by the language and library specification. A lock release followed by the appropriate acquisition, or a correctly paired release and acquire operation, can provide relevant synchronization. The distributed message model is an introduction, not a license to invent shared-memory guarantees. To justify a read of published data, trace a complete path from the writer action to the reader action through the synchronization operations required by the chosen API.
A total order can hide the missing guarantee
A logical clock can be constructed so that a happens-before relation implies increasing clock values. The reverse implication does not generally hold: ordered numbers can be assigned to events that remain causally unrelated. Breaking ties to produce a total order is useful for algorithms, but does not create a missing synchronization edge. During debugging, a nicely sorted timestamp log is therefore not a proof that shared data was safely published. Use an ordering diagram to display actual communication or synchronization edges, and distinguish that evidence from a convenient presentation order.