Realm.Builder.ha2 writes A2 into the recycled StringBuilder that
newResponse took from StringBuilderPool, but on the auth-int branch with
no precomputed entity-body hash it called StringUtils.toHexString, which
takes that same thread-local builder and resets it. The "POST:/secret:"
already written was discarded and A2 came out as the empty-body hash
twice, so the Digest response no longer bound the request method or the
target URI. Appending with appendBase16 keeps the hash in the buffer
already being built, which is what ha1 and newResponse already do for
HA1 and HA2. The two encoders emit identical lowercase, zero-padded hex,
so the digest is unchanged everywhere the branch was already correct.
Latent since #2148 replaced the EMPTY_ENTITY_MD5 constant with a
computed hash. The null-entityBodyHash branch is no longer reachable
from the request pipeline: #2276 wired setEntityBodyHash into
perRequestAuthorizationHeader and computeBodyHash never returns null, so
no wrong header reaches the wire today. It is still reached by the
nextnonce rotation in Interceptors and by Realms built through the
public API. The added RealmTest case checks the response against the
RFC 7616 A2 and fails on the current code.
Summary
This follows the entity-body-hash flow already used by the 401 and 407 interceptors. It does not modify Realm.java, so it remains compatible with #2269.
Verification
The full JDK 11 ./mvnw clean verify gate was not run because this environment provides JDK 21 only.
Codex on behalf of Pavel Ptashyts