GeezerVison

with a query registered in InMemoryPersistedQueryCache for an id, and a request that sends that id together with its own query text:

InMemoryPersistedQueryCacheTest > known queries take precedence over the query text in the execution input
Condition not satisfied:

printAstCompact(doc) == "{foo bar baz}"
|               |    |
{somethingElse} |    false

getPersistedQueryDocumentAsync reads executionInput.getQuery() first and only falls back to knownQueries when that text is null, empty or the APQ marker. so on a cache miss the request's text is what gets compiled and stored under the id, and every later caller that sends only the id is served that document. ApolloPersistedQuerySupport is not exposed because persistedQueryIdIsInvalid rejects text whose sha256 differs from the id, but the base PersistedQuerySupport returns false there, so with any other id scheme a client can replace a server registered query by sending its own text next to a known id. the custom cache in ApolloPersistedQuerySupportTest ("will use query hash in preference to query text") already treats the registered text as authoritative; the in memory cache did the opposite.

look the id up in knownQueries first and only register text from the execution input for an id that is not known. the apollo flow is unchanged, since text that passes the hash check is the registered text anyway.