A feed's sequence is the only ordering a consumer can trust, and here is the number that says so.
The measurement. On a session recorded from a real public push feed and committed here, in offset order, the wall clock steps backwards 547 times out of 2,024, by as much as 26 seconds, while the offset is monotone throughout. Each event also carries two clocks that disagree on 1,576 of 2,027, never in the other direction. Every rule here asserts on sequence continuity and none looks at a clock.
The four failures are injected, because the recorded session is clean and a consumer tested only against a healthy feed has been tested against nothing: a gap, an out-of-order delivery, a resend window and a correction for a record already consumed.
Two gap counts, and only one is true. An out-of-order delivery looks exactly like a gap on arrival and resolves when the offset behind it turns up. Alerting on the arrival-time count pages somebody for a feed that is working.
The offset lives in the sink, in the same transaction as the write. That claim is tested by failing the second write and asserting the first rolled back, which is the only crash a transaction is for. A version with the offset written outside the transaction passed every other test.
One breaking change, two settings. A required field with no default: BACKWARD rejects it with HTTP 409 before a message is written, NONE accepts it with 200 and every existing consumer meets a message it cannot decode.
The same normalisation twice, in Python and in a dependency-free Rust binary, compared line by line across all 2,027 events. Neither is the reference.
Licensing is a decision, not a footnote. Thirteen market data venues were checked and none permits a recorded session in a public repository. What is committed is a projection of a CC BY-SA feed with every free-text field dropped, because the offsets are facts and the comments are people's writing.
Verification. 8 required status checks, five of which re-run a measurement against a real broker, database, cache or compiler and fail if a byte of the result changed.