🟠 [Medium] isDGram cache write is an unsynchronized data race under WOLFSSL_RW_THREADED

Previously isDGramSock() was a pure read of the socket type. It now lazily writes dtlsCtx->isDGram and dtlsCtx->isDGramCached (bit-fields that share a storage unit with connected, and with processingPendingRecord when WOLFSSL_DTLS_CID is set). In a WOLFSSL_RW_THREADED build, EmbedSendTo (write thread) calls this on every send and EmbedReceiveFrom (read thread) calls it on the recvd == 0 path, so two threads can perform the first-time write concurrently. Writing adjacent bit-fields is a read-modify-write of the shared byte and is not atomic, so this is a data race (UB per the C memory model) on dtlsCtx. In practice it is benign because both threads write the same socket-type value and connected is not modified concurrently, but it can be reported by ThreadSanitizer CI and is a new race introduced by moving shared-state mutation into the hot I/O path. The rest of dtlsCtx peer state is already protected by peerLock precisely because these callbacks run concurrently.

Fix: Confirm intent for threaded builds; prefer computing the cache eagerly at fd assignment to remove the race from the data path entirely.

🔵 [Low] Single isDGram cache shared between rfd and wfd of potentially different socket types

The cache stores one isDGram flag per dtlsCtx, but it is consulted for both the read descriptor (dtlsCtx->rfd, via EmbedReceiveFrom) and the write descriptor (dtlsCtx->wfd, via EmbedSendTo). If an application sets rfd and wfd to different sockets of different types (legal via wolfSSL_set_read_fd/wolfSSL_set_write_fd), the first probe wins and the cached value is then incorrectly applied to the other descriptor. This is unusual for DTLS (rfd == wfd normally) and the cache is invalidated on each fd reassignment, so the common path is fine, but the original code re-probed the exact sfd passed in each time, whereas the cache is now keyed on the context rather than the descriptor.

Fix: Confirm the rfd==wfd (or same-type) assumption holds for all supported DTLS usages; document it next to the cache fields.

🔵 [Low] Confirm discarded explicit-nonce value is intentional across FIPS and epoch-order paths

The new code writes the record sequence number into the explicit nonce via PeekSEQ(ssl, epochOrder, args->iv) instead of wc_RNG_GenerateBlock. In every standard configuration this written value is subsequently overwritten before transmission: the non-FIPS/FIPSv2+ path copies the cipher-generated IV over it (XMEMCPY(out, ssl->encrypt.nonce + AESGCM_IMP_IV_SZ, ...) at line ~20741), and the legacy-FIPS path overwrites args->iv with ssl->keys.aead_exp_IV (line ~24839). So the real effect of this change is removing the per-record RNG draw; the seq value itself is intentionally discarded. Note also that the actual AAD seq is written with CUR_ORDER inside writeAeadAuthData (line ~20695) while PeekSEQ here uses epochOrder, so the two would differ for DTLS SCR PREV_ORDER retransmits, but this is harmless precisely because the PeekSEQ-written bytes are discarded. This is correct but subtle; a future reader could mistakenly assume the on-wire explicit nonce equals PeekSEQ.

Fix: Keep the change; add a short clarifying comment to prevent future misreading and confirm the RNG-removal is the intended win.

🔵 [Low] now_sec ignores clock_gettime failure, can return uninitialized time

clock_gettime()'s return value is cast to void and ts is not initialized, so on the (rare) failure of CLOCK_MONOTONIC the function returns garbage from the stack, corrupting all throughput/duration math. This is benchmark-only code so impact is limited, but it is trivial to harden.

Fix: Initialize ts and check the return value, returning a sentinel on failure.

🔵 [Low] -z (sink send) silently ignored when combined with -s (server)

The help text documents -z as client-only, but parse_args performs no validation when -z is passed together with -s. In server mode sinkSend is parsed and then never used, so a user invoking -s -z gets no warning that the flag had no effect. Minor usability nit in example code.

Fix: Emit a warning (or error) when -z is combined with -s so the no-op is visible to the user.