Blake2bpDigest buffers one 128-byte block per leaf: PARALLELISM_DEGREE is 4 (Blake2bpDigest.java:32) and buffer = new byte[512] (:39). update(byte[], int, int) measures the room left as 8*BLAKE2B_BLOCKBYTES - left (:66), 1024 less what is buffered. Once input is buffered, a call that does not fit goes wrong:

• at 1024 - left bytes or more, the flush copy at :70 ends past the buffer and throws ArrayIndexOutOfBoundsException;
• under 512 bytes, the flush is skipped and the tail copy at :101 throws the same way;
• in between, the leaf loop at :82-94 hashes the first 512 bytes of the new input ahead of the buffered bytes, a wrong digest with no exception.

update(byte) calls update(byte[], int, int) (:59), so a byte at a time fails on the 513th byte. One-shot hashing and messages of 512 bytes or fewer are unaffected.

Blake2spDigest has the same expression (Blake2spDigest.java:67) over eight 64-byte blocks (:31, :39), so it is consistent; there the buffer was short (#1363, fixed in 1.73). Here the limit is wrong: the BLAKE2 reference fills sizeof( S->buf ) - left of a 4 * BLAKE2B_BLOCKBYTES buffer (ref/blake2bp-ref.c:140, ref/blake2.h:84), and a 1024-byte buffer would still give wrong digests.

Reproduced on the released bcprov-jdk18on-1.86.jar with the new test's 1100-byte message (bytes 00 01 .. ff repeating): of the chunk sizes 1 to 1100, 946 threw ArrayIndexOutOfBoundsException (arraycopy: last destination index 513 out of bounds for byte[512] at size 1) and 76 (sizes 513 to 588) gave a wrong digest. Two calls of 10 and 600 bytes return the digest of M[10..522) || M[0..10) || M[522..610) rather than of M, and an org.bouncycastle.crypto.io.DigestOutputStream given write(buf, 0, 16) then write(buf, 16, 4096) throws with last destination index 1024. Current main and the 1.87-SNAPSHOT beta (1.87.0.20730) have the same Blake2bpDigest.java, and the beta behaves the same.

This change:

• measures the room left as PARALLELISM_DEGREE * BLAKE2B_BLOCKBYTES - left, the 512 bytes the buffer holds, as the leaf loop at :88 does;
• adds testMultiPartUpdate to Blake2bpDigestTest: an 1100-byte message hashed in one call against a fixed value, then in chunks of every size from 1 to 1100 and at every two-call split, unkeyed and keyed, and through update(byte), unkeyed; and
• registers Blake2bpDigestTest in crypto.test.AllTests. No suite referenced it and the Gradle test task runs only AllTest* classes (build.gradle:250), so it never ran. Blake2spDigestTest is in the same position (it passes) and is left out here.

The new test fails without the change (ArrayIndexOutOfBoundsException on the 513th one-byte update) and passes with it; the class's other tests and the other digest tests in crypto.test give the same results either way.

Base tree only: no META-INF/versions copy, module-info or OSGi change. A release-note entry is included, happy to move it to another block.