some fixes

data_acq_and_processing/processing/locator/src/payload_builder.cpp

@@ -95,6 +95,17 @@ constexpr std::uint32_t kMinTableColumns = 3U;  // [x_m, z_m, score, ...]
     return stream.str();
 }
 
+// Wall-clock epoch milliseconds at packet-build time. Unlike the human-readable
+// `tim` field (which has no date and rolls over at midnight), this is a
+// monotonically increasing generation stamp consumers can use to compute the
+// true age of a snapshot and detect a stalled pipeline.
+[[nodiscard]] auto epoch_millis_now() -> std::int64_t {
+    using Clock = std::chrono::system_clock;
+    return std::chrono::duration_cast<std::chrono::milliseconds>(
+        Clock::now().time_since_epoch()
+    ).count();
+}
+
 void append_u32_little_endian(std::vector<std::uint8_t>& buffer, std::uint32_t value) {
     buffer.push_back(static_cast<std::uint8_t>(value & 0xFFU));
     buffer.push_back(static_cast<std::uint8_t>((value >> 8U) & 0xFFU));
@@ -181,9 +192,15 @@ auto build_payload_json(
         });
     }
 
+    // Fix #48: stamp every packet with a numeric build-time generation
+    // (`gen`, epoch ms). A cached snapshot re-sent verbatim to a new client
+    // keeps its original `gen`, so a consumer can compute the snapshot's true
+    // age and flag a stalled pipeline (which `tim` alone cannot express, having
+    // no date and rolling over at midnight).
     const Json root{
         {"ver", protocol_version},
         {"tim", format_timestamp_now()},
+        {"gen", epoch_millis_now()},
         {"sts", status},
         {"obs", std::move(obs_array)},
     };

data_acq_and_processing/processing/locator/src/tcp_server.cpp

@@ -33,6 +33,14 @@ using Json = nlohmann::json;
 
 constexpr std::size_t kPacketHeaderSize = 8U;  // device_id u32 LE + payload_len u32 LE.
 
+// Per-recv()/send() socket timeout. Bounds how long a wedged peer can stall a
+// reader/writer loop and lets the reader re-check the stop flag periodically.
+constexpr long kSocketTimeoutSeconds = 5;
+
+// A socket-fed `vlc` speed older than this is considered stale and ignored, so
+// a client that disconnects (or stops sending) cannot pin a fixed speed forever.
+constexpr std::chrono::seconds kSocketSpeedStaleAfter{5};
+
 // Best-effort full-write helper: loops over write() until everything is sent
 // or an error occurs. Returns false on socket error or peer disconnect.
 [[nodiscard]] auto write_all(int socket_fd, const std::uint8_t* data, std::size_t size) -> bool {
@@ -48,6 +56,9 @@ constexpr std::size_t kPacketHeaderSize = 8U;  // device_id u32 LE + payload_len
             if (errno == EINTR) {
                 continue;
             }
+            // EAGAIN/EWOULDBLOCK here is a SO_SNDTIMEO send timeout: treat it as
+            // fatal so a stalled peer tears the session down rather than wedging
+            // the writer thread indefinitely.
             return false;
         }
         if (chunk == 0) {
@@ -58,15 +69,25 @@ constexpr std::size_t kPacketHeaderSize = 8U;  // device_id u32 LE + payload_len
     return true;
 }
 
-// Best-effort exact-read helper. Returns false if the peer closed the socket
-// or an unrecoverable error occurred before all bytes were read.
-[[nodiscard]] auto read_exact(int socket_fd, std::uint8_t* data, std::size_t size) -> bool {
+// Best-effort exact-read helper. Returns false if the peer closed the socket,
+// an unrecoverable error occurred, or stop was requested before all bytes were
+// read. A SO_RCVTIMEO timeout (EAGAIN/EWOULDBLOCK) is not fatal: we re-check the
+// stop flag and retry so a wedged peer cannot keep the reader thread alive.
+[[nodiscard]] auto read_exact(
+    int socket_fd,
+    std::uint8_t* data,
+    std::size_t size,
+    const std::atomic<bool>& stop_requested
+) -> bool {
     std::size_t consumed = 0;
     while (consumed < size) {
+        if (stop_requested.load(std::memory_order_acquire)) {
+            return false;
+        }
         const auto chunk = ::recv(socket_fd, data + consumed, size - consumed, 0);
         if (chunk < 0) {
-            if (errno == EINTR) {
-                continue;
+            if (errno == EINTR || errno == EAGAIN || errno == EWOULDBLOCK) {
+                continue;  // Interrupted or recv timeout: re-check stop and retry.
             }
             return false;
         }
@@ -107,6 +128,14 @@ void apply_socket_keepalive(int socket_fd) {
     int yes = 1;
     (void)::setsockopt(socket_fd, IPPROTO_TCP, TCP_NODELAY, &yes, sizeof(yes));
     (void)::setsockopt(socket_fd, SOL_SOCKET, SO_KEEPALIVE, &yes, sizeof(yes));
+    // Bound blocking recv()/send() so a wedged peer cannot stall a reader/writer
+    // thread forever. recv() timeouts are retried (stop-aware); a send() timeout
+    // is treated as a fatal write failure that tears the session down.
+    timeval timeout{};
+    timeout.tv_sec = kSocketTimeoutSeconds;
+    timeout.tv_usec = 0;
+    (void)::setsockopt(socket_fd, SOL_SOCKET, SO_RCVTIMEO, &timeout, sizeof(timeout));
+    (void)::setsockopt(socket_fd, SOL_SOCKET, SO_SNDTIMEO, &timeout, sizeof(timeout));
 }
 
 void shutdown_and_close(int& socket_fd) {
@@ -192,9 +221,14 @@ ClientSession::~ClientSession() {
     shutdown_and_close(socket_fd_);
 }
 
-void ClientSession::start(std::atomic<double>& shared_vlc_slot) {
+void ClientSession::start(
+    std::atomic<double>& shared_vlc_slot,
+    std::atomic<std::int64_t>& shared_vlc_at_ns
+) {
     writer_thread_ = std::thread([this]() { writer_loop(); });
-    reader_thread_ = std::thread([this, &shared_vlc_slot]() { reader_loop(shared_vlc_slot); });
+    reader_thread_ = std::thread([this, &shared_vlc_slot, &shared_vlc_at_ns]() {
+        reader_loop(shared_vlc_slot, shared_vlc_at_ns);
+    });
 }
 
 void ClientSession::enqueue(std::vector<std::uint8_t> packet) {
@@ -245,15 +279,24 @@ void ClientSession::writer_loop() {
         }
     }
     request_stop();
-    // Exited flag is set once both threads finish; reader_loop sets it.
+    // Mark exited only once BOTH loops have finished; otherwise a writer-only
+    // death (e.g. a send timeout while the reader still blocks on recv) would
+    // never be reaped. request_stop() above shuts the socket so the reader's
+    // recv() returns promptly.
+    if (loops_finished_.fetch_add(1, std::memory_order_acq_rel) + 1U == 2U) {
+        exited_.store(true, std::memory_order_release);
+    }
 }
 
-void ClientSession::reader_loop(std::atomic<double>& shared_vlc_slot) {
+void ClientSession::reader_loop(
+    std::atomic<double>& shared_vlc_slot,
+    std::atomic<std::int64_t>& shared_vlc_at_ns
+) {
     std::array<std::uint8_t, kPacketHeaderSize> header_buffer{};
     std::vector<std::uint8_t> payload_buffer;
 
     while (!stop_requested_.load(std::memory_order_acquire)) {
-        if (!read_exact(socket_fd_, header_buffer.data(), header_buffer.size())) {
+        if (!read_exact(socket_fd_, header_buffer.data(), header_buffer.size(), stop_requested_)) {
             break;
         }
         const auto payload_len = decode_u32_little_endian(header_buffer.data() + 4U);
@@ -266,7 +309,8 @@ void ClientSession::reader_loop(std::atomic<double>& shared_vlc_slot) {
         }
 
         payload_buffer.assign(payload_len, std::uint8_t{0});
-        if (payload_len > 0U && !read_exact(socket_fd_, payload_buffer.data(), payload_len)) {
+        if (payload_len > 0U
+            && !read_exact(socket_fd_, payload_buffer.data(), payload_len, stop_requested_)) {
             break;
         }
 
@@ -277,7 +321,13 @@ void ClientSession::reader_loop(std::atomic<double>& shared_vlc_slot) {
                 if (found != json.end() && found->is_number()) {
                     const double value = found->get<double>();
                     if (std::isfinite(value)) {
+                        // Stamp the value first, then the time, so a reader that
+                        // observes a fresh timestamp also observes the matching value.
                         shared_vlc_slot.store(value, std::memory_order_release);
+                        const auto now_ns = std::chrono::duration_cast<std::chrono::nanoseconds>(
+                            std::chrono::steady_clock::now().time_since_epoch()
+                        ).count();
+                        shared_vlc_at_ns.store(now_ns, std::memory_order_release);
                     }
                 }
             }
@@ -288,7 +338,11 @@ void ClientSession::reader_loop(std::atomic<double>& shared_vlc_slot) {
         }
     }
     request_stop();
-    exited_.store(true, std::memory_order_release);
+    // Mark exited only once BOTH loops have finished (see writer_loop), so a
+    // reader-only death still waits for the writer before this session is reaped.
+    if (loops_finished_.fetch_add(1, std::memory_order_acq_rel) + 1U == 2U) {
+        exited_.store(true, std::memory_order_release);
+    }
 }
 
 // ----- TcpServer ------------------------------------------------------------
@@ -409,11 +463,28 @@ void TcpServer::publish(const ipc::ResultCollection& collection, const FilterPar
     auto packet = encode_packet(payload_json, config_.device_id);
     cache_latest_packet(packet);
     broadcast_packet(packet);
+    // Reap exited sessions on the publish path too, so clients_ stays bounded
+    // even when no new connections arrive to trigger an accept-time reap. Uses a
+    // try-lock and joins outside the mutex, so publish() never blocks on it.
+    try_reap_finished_clients();
 }
 
 auto TcpServer::latest_socket_speed() const -> std::optional<double> {
+    // Read the timestamp first, then the value, mirroring the writer's order so a
+    // value seen here is at least as fresh as its timestamp.
+    const std::int64_t at_ns = latest_socket_speed_at_ns_.load(std::memory_order_acquire);
     const double value = latest_socket_speed_.load(std::memory_order_acquire);
-    if (std::isnan(value)) {
+    if (at_ns == 0 || std::isnan(value)) {
+        return std::nullopt;
+    }
+    // Expire stale socket-fed speeds: a client that stopped sending (or
+    // disconnected) must not pin a fixed speed indefinitely.
+    const auto now_ns = std::chrono::duration_cast<std::chrono::nanoseconds>(
+        std::chrono::steady_clock::now().time_since_epoch()
+    ).count();
+    const auto stale_after_ns =
+        std::chrono::duration_cast<std::chrono::nanoseconds>(kSocketSpeedStaleAfter).count();
+    if (now_ns - at_ns > stale_after_ns) {
         return std::nullopt;
     }
     return value;
@@ -457,7 +528,7 @@ void TcpServer::acceptor_loop() {
         if (snapshot.has_value()) {
             session->enqueue(*snapshot);
         }
-        session->start(latest_socket_speed_);
+        session->start(latest_socket_speed_, latest_socket_speed_at_ns_);
         enroll_client(std::move(session));
     }
 }
@@ -496,6 +567,35 @@ void TcpServer::reap_finished_clients() {
     }
 }
 
+void TcpServer::try_reap_finished_clients() {
+    std::vector<std::unique_ptr<ClientSession>> to_join;
+    {
+        // try_to_lock: if the acceptor already holds the mutex (it reaps too),
+        // skip this round rather than block the publish path.
+        std::unique_lock<std::mutex> guard(clients_mutex_, std::try_to_lock);
+        if (!guard.owns_lock()) {
+            return;
+        }
+        auto first_dead = std::partition(
+            clients_.begin(),
+            clients_.end(),
+            [](const std::unique_ptr<ClientSession>& session) {
+                return !session->has_exited();
+            }
+        );
+        for (auto it = first_dead; it != clients_.end(); ++it) {
+            to_join.push_back(std::move(*it));
+        }
+        clients_.erase(first_dead, clients_.end());
+    }
+    // Join outside the lock: a finished session's threads are done, so join()
+    // returns promptly without holding clients_mutex_.
+    for (auto& session : to_join) {
+        session->request_stop();
+        session->join();
+    }
+}
+
 void TcpServer::cache_latest_packet(std::vector<std::uint8_t> packet) {
     std::lock_guard<std::mutex> guard(latest_packet_mutex_);
     latest_packet_ = std::move(packet);