You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 92c17da
Browse filesBrowse the repository at this point in the historyBrowse files
🔑 Authenticate a retained answer against the wait it says it answers
A retained answer named a journal event and nothing checked what that event
was. The fixture proved how little that asks for: it pointed a consumed answer
at a Git effect and a pending one at a coroutine Close, carried no suspension
event at all, and the writer, its own read-back and both semantic tests took
it.
An answer is now held to the history it claims. The event it names has to be a
retained suspension request; that request has to be for this wait, by the
identity it carries twice over; the request is parsed rather than read, because
it reached the journal through a public durable operation; its fingerprint is
recomputed and compared, so a value judged against some other schema is not an
answer to this one; the value is judged again by the schema that wait retained;
and the row's state has to agree with the journal — consumed means the answer
was published, pending means it was not, and a publication may not precede its
request.
Every one of those questions is asked with the live contracts. The request
fingerprint had been written out twice already, once in delivery and once in
the resume that consumes an answer; it is now one function beside the request
parser, and all three callers share it. Judging the value is the same schema
compilation a delivery and an `<Elicit>` use.
Compiling that schema is an operation, so the semantic pass became one, and the
reader now closes its connection after reading content into memory rather than
holding it open through verification. That is what gate 6 always said, and the
gates after it never needed the file again.
The fixture carries two waits a run could really have reached — identifiers
derived from the run and the durable position, a request and its publication
for the answered one, a request alone for the wait the run is suspended at.
C8 gains the refusals that hold each rule: an event that is not a request, a
request for the other wait, a fingerprint from elsewhere, a value the schema
rejects, a consumed row with nothing published, and a pending row whose answer
already was.
Six of those seven pass against the check this replaces. That is what the gap
was.
Refs #602
0 commit comments