[#857] Iterate a snapshot of the connection set in AuthenticatedUsers.doPostResponse(modify)

[vharseko]

Fixes #857.

Problem

AuthenticatedUsers.doPostResponse(PostResponseModifyOperation) iterates the live per-user connection set (ConcurrentHashMap.newKeySet()), while updateAuthenticationInfo() → setAuthenticationInfo() re-registers each visited connection (remove + put) into that same set. The re-inserted node is appended to the tail of its hash-bin chain and the removed node keeps its next pointer, so the weakly-consistent iterator meets the connection again; with two or more connections in one bin the loop ping-pongs between them forever. The worker thread spins RUNNABLE and the bind that triggered the internal modify (password-policy last-login-time update) never gets a response — seen in CI as the PasswordPolicyTestCase.testResetWithLastLoginTime 600 s timeout.

Regression of #660: before it, the loop iterated a CopyOnWriteArraySet snapshot under the reentrant global write lock, so the nested remove/put could not affect the iteration. Both protections were removed together.

Fix

Iterate a snapshot (connectionSet.toArray(new ClientConnection[0])) instead of the live set — restores the copy-on-write iteration semantics while keeping registration lock-free. The delete/modifyDN paths are not affected: their sets are detached from the map first via removeSubtree.

Regression test

AuthenticatedUsersTestCase registers two connections with a constant hashCode() (forcing both into one bin of the connection set) under the same user DN and drives doPostResponse(modify) directly:

• without the fix the call never returns — deterministic reproduction, fails with the same ThreadTimeoutException ... 600000 signature as the CI hang;
• with the fix it passes in ~11 s: each connection is updated exactly once, stays registered, and caches the updated entry.