librats 0.3.1 → 0.5.1

package/native-src/src/dht.cpp

@@ -0,0 +1,2342 @@
+#include "dht.h"
+#include "network_utils.h"
+#include "logger.h"
+#include "socket.h"
+#include "json.hpp"
+#include <random>
+#include <algorithm>
+#include <sstream>
+#include <iomanip>
+#include <cstring>
+#include <cmath>
+#include <fstream>
+
+#ifdef _WIN32
+    #include <winsock2.h>
+    #include <ws2tcpip.h>
+#else
+    #include <arpa/inet.h>
+    #include <netinet/in.h>
+#endif
+
+// DHT module logging macros
+#define LOG_DHT_DEBUG(message) LOG_DEBUG("dht", message)
+#define LOG_DHT_INFO(message)  LOG_INFO("dht", message)
+#define LOG_DHT_WARN(message)  LOG_WARN("dht", message)
+#define LOG_DHT_ERROR(message) LOG_ERROR("dht", message)
+
+namespace librats {
+
+
+DhtClient::DhtClient(int port, const std::string& bind_address, const std::string& data_directory) 
+    : port_(port), bind_address_(bind_address), data_directory_(data_directory), 
+      socket_(INVALID_SOCKET_VALUE), running_(false) {
+    node_id_ = generate_node_id();
+    routing_table_.resize(NODE_ID_SIZE * 8);  // 160 buckets for 160-bit node IDs
+    
+    if (data_directory_.empty()) {
+        data_directory_ = ".";
+    }
+    
+    LOG_DHT_INFO("DHT client created with node ID: " << node_id_to_hex(node_id_) <<
+                 (bind_address_.empty() ? "" : " bind address: " + bind_address_) <<
+                 " data directory: " << data_directory_);
+}
+
+DhtClient::~DhtClient() {
+    stop();
+}
+
+bool DhtClient::start() {
+    if (running_) {
+        return true;
+    }
+    
+    LOG_DHT_INFO("Starting DHT client on port " << port_ <<
+                 (bind_address_.empty() ? "" : " bound to " + bind_address_));
+    
+    // Initialize socket library (safe to call multiple times)
+    if (!init_socket_library()) {
+        LOG_DHT_ERROR("Failed to initialize socket library");
+        return false;
+    }
+    
+    socket_ = create_udp_socket(port_, bind_address_);
+    if (!is_valid_socket(socket_)) {
+        LOG_DHT_ERROR("Failed to create dual-stack UDP socket");
+        return false;
+    }
+    
+    if (!set_socket_nonblocking(socket_)) {
+        LOG_DHT_WARN("Failed to set socket to non-blocking mode");
+    }
+    
+    running_ = true;
+    
+    // Load saved routing table before starting threads
+    if (load_routing_table()) {
+        LOG_DHT_INFO("Loaded routing table from disk (" << get_routing_table_size() << " nodes)");
+    }
+    
+    // Start network and maintenance threads
+    network_thread_ = std::thread(&DhtClient::network_loop, this);
+    maintenance_thread_ = std::thread(&DhtClient::maintenance_loop, this);
+    
+    LOG_DHT_INFO("DHT client started successfully");
+    return true;
+}
+
+void DhtClient::stop() {
+    if (!running_) {
+        return;
+    }
+    
+    LOG_DHT_INFO("Stopping DHT client");
+    
+    // Trigger immediate shutdown of all background threads
+    shutdown_immediate();
+    
+    // Wait for threads to finish
+    if (network_thread_.joinable()) {
+        network_thread_.join();
+    }
+    if (maintenance_thread_.joinable()) {
+        maintenance_thread_.join();
+    }
+    
+    // Save routing table before closing
+    if (save_routing_table()) {
+        LOG_DHT_INFO("Saved routing table to disk (" << get_routing_table_size() << " nodes)");
+    }
+    
+    // Close socket
+    if (is_valid_socket(socket_)) {
+        close_socket(socket_);
+        socket_ = INVALID_SOCKET_VALUE;
+    }
+    
+    LOG_DHT_INFO("DHT client stopped");
+}
+
+void DhtClient::shutdown_immediate() {
+    LOG_DHT_INFO("Triggering immediate shutdown of DHT background threads");
+    
+    running_.store(false);
+    
+    // Notify all waiting threads to wake up immediately
+    shutdown_cv_.notify_all();
+}
+
+bool DhtClient::bootstrap(const std::vector<Peer>& bootstrap_nodes) {
+    if (!running_) {
+        LOG_DHT_ERROR("DHT client not running");
+        return false;
+    }
+    
+    LOG_DHT_INFO("Bootstrapping DHT with " << bootstrap_nodes.size() << " nodes");
+    LOG_DHT_DEBUG("Bootstrap nodes:");
+    for (const auto& peer : bootstrap_nodes) {
+        LOG_DHT_DEBUG("  - " << peer.ip << ":" << peer.port);
+    }
+    
+
+    
+    // Send ping to bootstrap nodes
+    LOG_DHT_DEBUG("Sending PING to all bootstrap nodes");
+    for (const auto& peer : bootstrap_nodes) {
+        send_krpc_ping(peer);
+    }
+    
+    // Start node discovery by finding our own node
+    LOG_DHT_DEBUG("Starting node discovery by finding our own node ID: " << node_id_to_hex(node_id_));
+    for (const auto& peer : bootstrap_nodes) {
+        send_krpc_find_node(peer, node_id_);
+    }
+    
+    LOG_DHT_DEBUG("Bootstrap process initiated");
+    return true;
+}
+
+bool DhtClient::find_peers(const InfoHash& info_hash, PeerDiscoveryCallback callback) {
+    if (!running_) {
+        LOG_DHT_ERROR("DHT client not running");
+        return false;
+    }
+    
+    std::string hash_key = node_id_to_hex(info_hash);
+    LOG_DHT_INFO("Finding peers for info hash: " << hash_key);
+    
+    // Get initial nodes from routing table
+    auto closest_nodes = find_closest_nodes(info_hash, K_BUCKET_SIZE);
+    
+    if (closest_nodes.empty()) {
+        LOG_DHT_WARN("No nodes in routing table to query for info_hash " << hash_key);
+        return false;
+    }
+    
+    DeferredCallbacks deferred;
+    
+    {
+        std::lock_guard<std::mutex> lock(pending_searches_mutex_);
+        
+        // Check if a search is already ongoing for this info_hash
+        auto search_it = pending_searches_.find(hash_key);
+        if (search_it != pending_searches_.end()) {
+            // Search already in progress - just add the callback to the list
+            LOG_DHT_INFO("Search already in progress for info hash " << hash_key << " - adding callback to existing search");
+            search_it->second.callbacks.push_back(callback);
+            return true;
+        }
+        
+        // Create new search
+        PendingSearch new_search(info_hash);
+        new_search.callbacks.push_back(callback);
+        
+        // Initialize search_nodes with closest nodes from routing table (already sorted)
+        new_search.search_nodes = std::move(closest_nodes);
+        
+        auto insert_result = pending_searches_.emplace(hash_key, std::move(new_search));
+        PendingSearch& search_ref = insert_result.first->second;
+        
+        LOG_DHT_DEBUG("Initialized search with " << search_ref.search_nodes.size() << " nodes from routing table");
+        
+        // Start sending requests
+        add_search_requests(search_ref, deferred);
+    }
+    
+    // Invoke callbacks outside the lock to avoid deadlock
+    deferred.invoke();
+    
+    return true;
+}
+
+bool DhtClient::announce_peer(const InfoHash& info_hash, uint16_t port) {
+    if (!running_) {
+        LOG_DHT_ERROR("DHT client not running");
+        return false;
+    }
+    
+    if (port == 0) {
+        port = port_;
+    }
+    
+    LOG_DHT_INFO("Announcing peer for info hash: " << node_id_to_hex(info_hash) << " on port " << port);
+    
+    // First find nodes close to the info hash and send get_peers to them
+    // This is the proper BEP 5 flow: get_peers -> collect tokens -> announce_peer
+    auto closest_nodes = find_closest_nodes(info_hash, ALPHA);
+    for (const auto& node : closest_nodes) {
+            // Generate transaction ID and track this as a pending announce for KRPC
+            std::string transaction_id = KrpcProtocol::generate_transaction_id();
+            
+            {
+                std::lock_guard<std::mutex> lock(pending_announces_mutex_);
+                pending_announces_.emplace(transaction_id, PendingAnnounce(info_hash, port));
+            }
+            
+            auto message = KrpcProtocol::create_get_peers_query(transaction_id, node_id_, info_hash);
+            send_krpc_message(message, node.peer);
+    }
+    
+    return true;
+}
+
+size_t DhtClient::get_routing_table_size() const {
+    std::lock_guard<std::mutex> lock(routing_table_mutex_);
+    size_t total = 0;
+    for (const auto& bucket : routing_table_) {
+        total += bucket.size();
+    }
+    return total;
+}
+
+size_t DhtClient::get_pending_ping_verifications_count() const {
+    std::lock_guard<std::mutex> lock(pending_pings_mutex_);
+    return pending_pings_.size();
+}
+
+bool DhtClient::is_search_active(const InfoHash& info_hash) const {
+    std::lock_guard<std::mutex> lock(pending_searches_mutex_);
+    std::string hash_key = node_id_to_hex(info_hash);
+    auto it = pending_searches_.find(hash_key);
+    return it != pending_searches_.end() && !it->second.is_finished;
+}
+
+size_t DhtClient::get_active_searches_count() const {
+    std::lock_guard<std::mutex> lock(pending_searches_mutex_);
+    size_t count = 0;
+    for (const auto& [hash, search] : pending_searches_) {
+        if (!search.is_finished) {
+            count++;
+        }
+    }
+    return count;
+}
+
+std::vector<Peer> DhtClient::get_default_bootstrap_nodes() {
+    return {
+        {"router.bittorrent.com", 6881},
+        {"dht.transmissionbt.com", 6881},
+        {"router.utorrent.com", 6881},
+        {"dht.aelitis.com", 6881}
+    };
+}
+
+void DhtClient::network_loop() {
+    LOG_DHT_DEBUG("Network loop started");
+    
+    while (running_) {
+        Peer sender;
+        auto data = receive_udp_data(socket_, 1500, sender);  // MTU size
+        
+        if (!data.empty()) {
+            LOG_DHT_DEBUG("Received " << data.size() << " bytes from " << sender.ip << ":" << sender.port);
+            handle_message(data, sender);
+        }
+        
+        // Use conditional variable for responsive shutdown
+        {
+            std::unique_lock<std::mutex> lock(shutdown_mutex_);
+            if (shutdown_cv_.wait_for(lock, std::chrono::milliseconds(10), [this] { return !running_.load(); })) {
+                break;
+            }
+        }
+    }
+    
+    LOG_DHT_DEBUG("Network loop stopped");
+}
+
+void DhtClient::maintenance_loop() {
+    LOG_DHT_DEBUG("Maintenance loop started");
+    
+    auto last_bucket_refresh = std::chrono::steady_clock::now();
+    auto last_ping_verification_cleanup = std::chrono::steady_clock::now();
+    auto last_general_cleanup = std::chrono::steady_clock::now();
+    auto last_stats_print = std::chrono::steady_clock::now();
+    auto last_search_timeout_check = std::chrono::steady_clock::now();
+    auto last_search_node_cleanup = std::chrono::steady_clock::now();
+    auto last_routing_table_save = std::chrono::steady_clock::now();
+    
+    while (running_) {
+        auto now = std::chrono::steady_clock::now();
+
+        // Check for timed out search requests every 2 seconds (frequent check)
+        if (now - last_search_timeout_check >= std::chrono::seconds(2)) {
+            cleanup_timed_out_search_requests();
+            last_search_timeout_check = now;
+        }
+        
+        // Clean up finalized node_states entries in active searches every 10 seconds
+        if (now - last_search_node_cleanup >= std::chrono::seconds(10)) {
+            cleanup_search_node_states();
+            last_search_node_cleanup = now;
+        }
+        
+        // General cleanup operations every 1 minute (like previously)
+        if (now - last_general_cleanup >= std::chrono::minutes(1)) {
+            // Cleanup stale nodes every 1 minute
+            cleanup_stale_nodes();
+            
+            // Cleanup stale peer tokens
+            cleanup_stale_peer_tokens();
+            
+            // Cleanup stale pending announces
+            cleanup_stale_announces();
+            
+            // Cleanup stale pending searches
+            cleanup_stale_searches();
+            
+            // Cleanup stale announced peers
+            cleanup_stale_announced_peers();
+            
+            last_general_cleanup = now;
+        }
+        
+        // Refresh buckets every 30 minutes
+        if (now - last_bucket_refresh >= std::chrono::minutes(30)) {
+            refresh_buckets();
+            last_bucket_refresh = now;
+        }
+        
+        // Frequent maintenance: ping verifications time out at ~30s, so check often
+        if (now - last_ping_verification_cleanup >= std::chrono::seconds(30)) {
+            cleanup_stale_ping_verifications();
+            last_ping_verification_cleanup = now;
+        }
+        
+        // Print DHT statistics every 10 seconds
+        if (now - last_stats_print >= std::chrono::seconds(10)) {
+            print_statistics();
+            last_stats_print = now;
+        }
+        
+        // Save routing table every 5 minutes
+        if (now - last_routing_table_save >= std::chrono::minutes(5)) {
+            if (save_routing_table()) {
+                LOG_DHT_DEBUG("Periodic routing table save completed");
+            }
+            last_routing_table_save = now;
+        }
+        
+        // Execute maintenance loop every 1 second
+        {
+            std::unique_lock<std::mutex> lock(shutdown_mutex_);
+            if (shutdown_cv_.wait_for(lock, std::chrono::seconds(1), [this] { return !running_.load(); })) {
+                break;
+            }
+        }
+    }
+    
+    LOG_DHT_DEBUG("Maintenance loop stopped");
+}
+
+void DhtClient::handle_message(const std::vector<uint8_t>& data, const Peer& sender) {
+    LOG_DHT_DEBUG("Processing message of " << data.size() << " bytes from " << sender.ip << ":" << sender.port);
+    
+    auto krpc_message = KrpcProtocol::decode_message(data);
+    if (!krpc_message) {
+        LOG_DHT_WARN("Failed to decode KRPC message from " << sender.ip << ":" << sender.port);
+        return;
+    }
+    
+    handle_krpc_message(*krpc_message, sender);
+}
+
+void DhtClient::add_node(const DhtNode& node, bool confirmed) {
+    std::lock_guard<std::mutex> ping_lock(pending_pings_mutex_);
+    std::lock_guard<std::mutex> lock(routing_table_mutex_);
+
+    int bucket_index = get_bucket_index(node.id);
+    auto& bucket = routing_table_[bucket_index];
+
+    LOG_DHT_DEBUG("Adding node " << node_id_to_hex(node.id) << " at " << node.peer.ip << ":" << node.peer.port 
+                  << " to bucket " << bucket_index << " (confirmed=" << confirmed << ")");
+        
+    // Check if node already exists
+    auto it = std::find_if(bucket.begin(), bucket.end(),
+                          [&node](const DhtNode& existing) {
+                              return existing.id == node.id;
+                          });
+
+    if (it != bucket.end()) {
+        // Update existing node - mark as successful since it contacted us
+        LOG_DHT_DEBUG("Node " << node_id_to_hex(node.id) << " already exists in bucket " << bucket_index << ", updating");
+        it->peer = node.peer;
+        if (confirmed) {
+            it->mark_success();
+        }
+        return;
+    }
+
+    // Bucket has space - just add
+    if (bucket.size() < K_BUCKET_SIZE) {
+        DhtNode new_node = node;
+        if (confirmed) {
+            new_node.fail_count = 0;  // Node contacted us, so it's confirmed good
+        }
+        bucket.push_back(new_node);
+        LOG_DHT_DEBUG("Added new node " << node_id_to_hex(node.id) << " to bucket " << bucket_index << " (size: " << bucket.size() << "/" << K_BUCKET_SIZE << ")");
+        return;
+    }
+
+    // Bucket is full - first check for nodes with failures (stale nodes)
+    auto worst_it = std::max_element(bucket.begin(), bucket.end(),
+        [](const DhtNode& a, const DhtNode& b) {
+            // Find node with highest fail_count
+            return a.fail_count < b.fail_count;
+        });
+    
+    if (worst_it != bucket.end() && worst_it->fail_count > 0) {
+        // Found a stale node - replace it immediately
+        LOG_DHT_DEBUG("Replacing stale node " << node_id_to_hex(worst_it->id) 
+                      << " (fail_count=" << static_cast<int>(worst_it->fail_count) << ")"
+                      << " with " << node_id_to_hex(node.id));
+        DhtNode new_node = node;
+        if (confirmed) {
+            new_node.fail_count = 0;  // Node contacted us, so it's confirmed good
+        }
+        // else: keep fail_count = 0xff (unpinged) from constructor
+        *worst_it = new_node;
+        return;
+    }
+
+    // All nodes are good - find the "worst" good node for ping verification
+    // Worst = highest RTT among nodes not already being pinged
+    DhtNode* worst = nullptr;
+    for (auto& existing : bucket) {
+        if (nodes_being_replaced_.find(existing.id) == nodes_being_replaced_.end()) {
+            if (!worst || existing.is_worse_than(*worst)) {
+                worst = &existing;
+            }
+        }
+    }
+    
+    if (!worst) {
+        LOG_DHT_DEBUG("All nodes in bucket already have pending pings - dropping candidate " << node_id_to_hex(node.id));
+        return;
+    }
+    
+    // Initiate ping to worst node - if it doesn't respond, replace with candidate
+    LOG_DHT_DEBUG("All nodes good, pinging worst node " << node_id_to_hex(worst->id) 
+                  << " (rtt=" << worst->rtt << "ms) to verify");
+    initiate_ping_verification(node, *worst, bucket_index);
+}
+
+std::vector<DhtNode> DhtClient::find_closest_nodes(const NodeId& target, size_t count) {
+    std::lock_guard<std::mutex> lock(routing_table_mutex_);
+    
+    auto result = find_closest_nodes_unlocked(target, count);
+    
+    return result;
+}
+
+std::vector<DhtNode> DhtClient::find_closest_nodes_unlocked(const NodeId& target, size_t count) {
+    LOG_DHT_DEBUG("Finding closest nodes to target " << node_id_to_hex(target) << " (max " << count << " nodes)");
+    
+    // Find closest bucket to target
+    int target_bucket = get_bucket_index(target);
+    
+    // Candidate nodes to be closest to target
+    std::vector<DhtNode> candidates;
+    // Reserve extra space: 3x count + buffer for 2 full buckets to avoid reallocation
+    candidates.reserve(count * 3 + K_BUCKET_SIZE * 2);
+    
+    // Add nodes from ideal bucket
+    if (target_bucket < routing_table_.size()) {
+        const auto& bucket = routing_table_[target_bucket];
+        candidates.insert(candidates.end(), bucket.begin(), bucket.end());
+        LOG_DHT_DEBUG("Collected " << bucket.size() << " nodes from target bucket " << target_bucket);
+    }
+    
+    // Add nodes from buckets above and below the ideal bucket
+    // Collect more candidates than needed to ensure we get the actual closest ones after sorting
+    size_t desired_candidates = count * 3;  // Collect 3x more candidates for better selection
+    int low = target_bucket - 1;
+    int high = target_bucket + 1;
+    const int max_bucket_index = static_cast<int>(routing_table_.size()) - 1;
+    int buckets_checked = 1;  // Already checked target_bucket
+    
+    while (candidates.size() < desired_candidates && (low >= 0 || high <= max_bucket_index)) {
+        // Search left (closer buckets)
+        if (low >= 0) {
+            const auto& bucket = routing_table_[low];
+            if (!bucket.empty()) {
+                candidates.insert(candidates.end(), bucket.begin(), bucket.end());
+                LOG_DHT_DEBUG("Collected " << bucket.size() << " nodes from bucket " << low);
+            }
+            low--;
+            buckets_checked++;
+        }
+        
+        // Search right (farther buckets)
+        if (high <= max_bucket_index) {
+            const auto& bucket = routing_table_[high];
+            if (!bucket.empty()) {
+                candidates.insert(candidates.end(), bucket.begin(), bucket.end());
+                LOG_DHT_DEBUG("Collected " << bucket.size() << " nodes from bucket " << high);
+            }
+            high++;
+            buckets_checked++;
+        }
+    }
+    
+    LOG_DHT_DEBUG("Bucket-aware collection: checked " << buckets_checked << " buckets, collected " 
+                  << candidates.size() << " candidate nodes around target bucket " << target_bucket);
+    
+    if (candidates.empty()) {
+        LOG_DHT_DEBUG("No candidates found in routing table");
+        return candidates;
+    }
+    
+    // Use partial_sort to efficiently get only the 'count' closest nodes - O(n log k) vs O(n log n)
+    size_t sort_count = (std::min)(count, candidates.size());
+    std::partial_sort(
+        candidates.begin(), 
+        candidates.begin() + sort_count, 
+        candidates.end(),
+        [&target, this](const DhtNode& a, const DhtNode& b) {
+            return is_closer(a.id, b.id, target);
+        }
+    );
+    
+    // Return up to 'count' closest nodes
+    if (candidates.size() > count) {
+        candidates.resize(count);
+    }
+    
+    LOG_DHT_DEBUG("Found " << candidates.size() << " closest nodes to target " << node_id_to_hex(target));
+    for (size_t i = 0; i < candidates.size(); ++i) {
+        LOG_DHT_DEBUG("  [" << i << "] " << node_id_to_hex(candidates[i].id) << " at " << candidates[i].peer.ip << ":" << candidates[i].peer.port);
+    }
+
+    // Debug alternative: Compare with full routing table algorithm
+    /*
+    candidates.clear();
+    for (const auto& bucket : routing_table_) {
+        candidates.insert(candidates.end(), bucket.begin(), bucket.end());
+    }
+    sort_count = (std::min)(count, candidates.size());
+    std::partial_sort(
+        candidates.begin(),
+        candidates.begin() + sort_count,
+        candidates.end(),
+        [&target, this](const DhtNode& a, const DhtNode& b) {
+            return is_closer(a.id, b.id, target);
+        }
+    );
+    // Return up to 'count' closest nodes
+    if (candidates.size() > count) {
+        candidates.resize(count);
+    }
+    LOG_DHT_DEBUG("Found " << candidates.size() << " closest nodes to target " << node_id_to_hex(target));
+    for (size_t i = 0; i < candidates.size(); ++i) {
+        LOG_DHT_DEBUG("  +[" << i << "] " << node_id_to_hex(candidates[i].id) << " at " << candidates[i].peer.ip << ":" << candidates[i].peer.port);
+    }
+    */
+    // End of debug alternative
+    
+    return candidates;
+}
+
+int DhtClient::get_bucket_index(const NodeId& id) {
+    NodeId distance = xor_distance(node_id_, id);
+    
+    // Find the position of the most significant bit
+    for (int i = 0; i < NODE_ID_SIZE; ++i) {
+        if (distance[i] != 0) {
+            for (int j = 7; j >= 0; --j) {
+                if (distance[i] & (1 << j)) {
+                    return i * 8 + (7 - j);
+                }
+            }
+        }
+    }
+    
+    return NODE_ID_SIZE * 8 - 1;  // All bits are 0, maximum distance
+}
+
+
+
+// KRPC message handling
+void DhtClient::handle_krpc_message(const KrpcMessage& message, const Peer& sender) {
+    LOG_DHT_DEBUG("Handling KRPC message type " << static_cast<int>(message.type) << " from " << sender.ip << ":" << sender.port);
+    
+    switch (message.type) {
+        case KrpcMessageType::Query:
+            switch (message.query_type) {
+                case KrpcQueryType::Ping:
+                    handle_krpc_ping(message, sender);
+                    break;
+                case KrpcQueryType::FindNode:
+                    handle_krpc_find_node(message, sender);
+                    break;
+                case KrpcQueryType::GetPeers:
+                    handle_krpc_get_peers(message, sender);
+                    break;
+                case KrpcQueryType::AnnouncePeer:
+                    handle_krpc_announce_peer(message, sender);
+                    break;
+            }
+            break;
+        case KrpcMessageType::Response:
+            handle_krpc_response(message, sender);
+            break;
+        case KrpcMessageType::Error:
+            handle_krpc_error(message, sender);
+            break;
+    }
+}
+
+void DhtClient::handle_krpc_ping(const KrpcMessage& message, const Peer& sender) {
+    LOG_DHT_DEBUG("Handling KRPC PING from " << node_id_to_hex(message.sender_id) << " at " << sender.ip << ":" << sender.port);
+    
+    // Add sender to routing table
+    KrpcNode krpc_node(message.sender_id, sender.ip, sender.port);
+    DhtNode sender_node = krpc_node_to_dht_node(krpc_node);
+    add_node(sender_node);
+    
+    // Respond with ping response
+    auto response = KrpcProtocol::create_ping_response(message.transaction_id, node_id_);
+    send_krpc_message(response, sender);
+}
+
+void DhtClient::handle_krpc_find_node(const KrpcMessage& message, const Peer& sender) {
+    LOG_DHT_DEBUG("Handling KRPC FIND_NODE from " << node_id_to_hex(message.sender_id) << " at " << sender.ip << ":" << sender.port);
+    
+    // Add sender to routing table
+    KrpcNode krpc_node(message.sender_id, sender.ip, sender.port);
+    DhtNode sender_node = krpc_node_to_dht_node(krpc_node);
+    add_node(sender_node);
+    
+    // Find closest nodes
+    auto closest_nodes = find_closest_nodes(message.target_id, K_BUCKET_SIZE);
+    auto krpc_nodes = dht_nodes_to_krpc_nodes(closest_nodes);
+    
+    // Respond with closest nodes
+    auto response = KrpcProtocol::create_find_node_response(message.transaction_id, node_id_, krpc_nodes);
+    send_krpc_message(response, sender);
+}
+
+void DhtClient::handle_krpc_get_peers(const KrpcMessage& message, const Peer& sender) {
+    LOG_DHT_DEBUG("Handling KRPC GET_PEERS from " << node_id_to_hex(message.sender_id) << " at " << sender.ip << ":" << sender.port << " for info_hash " << node_id_to_hex(message.info_hash));
+    
+    // Add sender to routing table
+    KrpcNode krpc_node(message.sender_id, sender.ip, sender.port);
+    DhtNode sender_node = krpc_node_to_dht_node(krpc_node);
+    add_node(sender_node);
+    
+    // Generate a token for this peer
+    std::string token = generate_token(sender);
+    
+    // First check if we have announced peers for this info_hash
+    auto announced_peers = get_announced_peers(message.info_hash);
+    
+    KrpcMessage response;
+    if (!announced_peers.empty()) {
+        // Return the peers we have stored
+        response = KrpcProtocol::create_get_peers_response(message.transaction_id, node_id_, announced_peers, token);
+        LOG_DHT_DEBUG("Responding to KRPC GET_PEERS with " << announced_peers.size() << " announced peers for info_hash " << node_id_to_hex(message.info_hash));
+    } else {
+        // Return closest nodes
+        auto closest_nodes = find_closest_nodes(message.info_hash, K_BUCKET_SIZE);
+        auto krpc_nodes = dht_nodes_to_krpc_nodes(closest_nodes);
+        response = KrpcProtocol::create_get_peers_response_with_nodes(message.transaction_id, node_id_, krpc_nodes, token);
+        LOG_DHT_DEBUG("Responding to KRPC GET_PEERS with " << krpc_nodes.size() << " closest nodes for info_hash " << node_id_to_hex(message.info_hash));
+    }
+    
+    send_krpc_message(response, sender);
+}
+
+void DhtClient::handle_krpc_announce_peer(const KrpcMessage& message, const Peer& sender) {
+    LOG_DHT_DEBUG("Handling KRPC ANNOUNCE_PEER from " << node_id_to_hex(message.sender_id) << " at " << sender.ip << ":" << sender.port);
+    
+    // Verify token
+    if (!verify_token(sender, message.token)) {
+        LOG_DHT_WARN("Invalid token from " << sender.ip << ":" << sender.port << " for KRPC ANNOUNCE_PEER");
+        auto error = KrpcProtocol::create_error(message.transaction_id, KrpcErrorCode::ProtocolError, "Invalid token");
+        send_krpc_message(error, sender);
+        return;
+    }
+    
+    // Add sender to routing table
+    KrpcNode krpc_node(message.sender_id, sender.ip, sender.port);
+    DhtNode sender_node = krpc_node_to_dht_node(krpc_node);
+    add_node(sender_node);
+    
+    // Store the peer announcement
+    Peer announcing_peer(sender.ip, message.port);
+    store_announced_peer(message.info_hash, announcing_peer);
+    
+    // Respond with acknowledgment
+    auto response = KrpcProtocol::create_announce_peer_response(message.transaction_id, node_id_);
+    send_krpc_message(response, sender);
+}
+
+void DhtClient::handle_krpc_response(const KrpcMessage& message, const Peer& sender) {
+    LOG_DHT_DEBUG("Handling KRPC response from " << sender.ip << ":" << sender.port);
+    
+    // Check if this is a ping verification response before normal processing
+    handle_ping_verification_response(message.transaction_id, message.response_id, sender);
+    
+    // Add responder to routing table
+    KrpcNode krpc_node(message.response_id, sender.ip, sender.port);
+    DhtNode sender_node = krpc_node_to_dht_node(krpc_node);
+    add_node(sender_node);
+    
+    // Add any nodes from the response (these are nodes we heard about, not confirmed)
+    for (const auto& node : message.nodes) {
+        DhtNode dht_node = krpc_node_to_dht_node(node);
+        add_node(dht_node, false);  // Not confirmed - just heard about from another node
+    }
+    
+    // Check if this is a response to a pending search (get_peers with peers)
+    if (!message.peers.empty()) {
+        handle_get_peers_response_for_search(message.transaction_id, sender, message.peers);
+    }
+    // Check if this is a response to a pending search (get_peers with nodes)
+    else if (!message.nodes.empty()) {
+        handle_get_peers_response_with_nodes(message.transaction_id, sender, message.nodes);
+    }
+    else {
+        // Empty response (no peers, no nodes) - still need to mark as responded
+        // This can happen when a node has no information about the info_hash
+        handle_get_peers_empty_response(message.transaction_id, sender);
+    }
+    
+    // Check if this is a response to a pending announce (get_peers with token)
+    if (!message.token.empty()) {
+        handle_get_peers_response_for_announce(message.transaction_id, sender, message.token);
+    }
+    
+    // Clean up finished searches AFTER all response data has been processed
+    // This ensures peers and nodes are fully handled before removing the search
+    {
+        std::lock_guard<std::mutex> lock(pending_searches_mutex_);
+        auto trans_it = transaction_to_search_.find(message.transaction_id);
+        if (trans_it != transaction_to_search_.end()) {
+            const std::string& hash_key = trans_it->second.info_hash_hex;
+            auto search_it = pending_searches_.find(hash_key);
+            if (search_it != pending_searches_.end() && search_it->second.is_finished) {
+                LOG_DHT_DEBUG("Cleaning up finished search for info_hash " << hash_key 
+                              << " after processing transaction " << message.transaction_id);
+                pending_searches_.erase(search_it);
+            }
+            // Always remove the transaction mapping after processing
+            transaction_to_search_.erase(trans_it);
+        }
+    }
+}
+
+void DhtClient::handle_krpc_error(const KrpcMessage& message, const Peer& sender) {
+    LOG_DHT_WARN("Received KRPC error from " << sender.ip << ":" << sender.port 
+                 << " - Code: " << static_cast<int>(message.error_code) 
+                 << " Message: " << message.error_message);
+}
+
+// KRPC sending functions
+bool DhtClient::send_krpc_message(const KrpcMessage& message, const Peer& peer) {
+    auto data = KrpcProtocol::encode_message(message);
+    if (data.empty()) {
+        LOG_DHT_ERROR("Failed to encode KRPC message");
+        return false;
+    }
+    
+    LOG_DHT_DEBUG("Sending KRPC message (" << data.size() << " bytes) to " << peer.ip << ":" << peer.port);
+    int result = send_udp_data(socket_, data, peer);
+    
+    if (result > 0) {
+        LOG_DHT_DEBUG("Successfully sent KRPC message to " << peer.ip << ":" << peer.port);
+    } else {
+        LOG_DHT_ERROR("Failed to send KRPC message to " << peer.ip << ":" << peer.port);
+    }
+    
+    return result > 0;
+}
+
+void DhtClient::send_krpc_ping(const Peer& peer) {
+    std::string transaction_id = KrpcProtocol::generate_transaction_id();
+    auto message = KrpcProtocol::create_ping_query(transaction_id, node_id_);
+    send_krpc_message(message, peer);
+}
+
+void DhtClient::send_krpc_find_node(const Peer& peer, const NodeId& target) {
+    std::string transaction_id = KrpcProtocol::generate_transaction_id();
+    auto message = KrpcProtocol::create_find_node_query(transaction_id, node_id_, target);
+    send_krpc_message(message, peer);
+}
+
+void DhtClient::send_krpc_get_peers(const Peer& peer, const InfoHash& info_hash) {
+    std::string transaction_id = KrpcProtocol::generate_transaction_id();
+    auto message = KrpcProtocol::create_get_peers_query(transaction_id, node_id_, info_hash);
+    send_krpc_message(message, peer);
+}
+
+void DhtClient::send_krpc_announce_peer(const Peer& peer, const InfoHash& info_hash, uint16_t port, const std::string& token) {
+    std::string transaction_id = KrpcProtocol::generate_transaction_id();
+    auto message = KrpcProtocol::create_announce_peer_query(transaction_id, node_id_, info_hash, port, token);
+    send_krpc_message(message, peer);
+}
+
+// Conversion utilities
+KrpcNode DhtClient::dht_node_to_krpc_node(const DhtNode& node) {
+    return KrpcNode(node.id, node.peer.ip, node.peer.port);
+}
+
+DhtNode DhtClient::krpc_node_to_dht_node(const KrpcNode& node) {
+    Peer peer(node.ip, node.port);
+    return DhtNode(node.id, peer);
+}
+
+std::vector<KrpcNode> DhtClient::dht_nodes_to_krpc_nodes(const std::vector<DhtNode>& nodes) {
+    std::vector<KrpcNode> krpc_nodes;
+    krpc_nodes.reserve(nodes.size());
+    for (const auto& node : nodes) {
+        krpc_nodes.push_back(dht_node_to_krpc_node(node));
+    }
+    return krpc_nodes;
+}
+
+std::vector<DhtNode> DhtClient::krpc_nodes_to_dht_nodes(const std::vector<KrpcNode>& nodes) {
+    std::vector<DhtNode> dht_nodes;
+    dht_nodes.reserve(nodes.size());
+    for (const auto& node : nodes) {
+        dht_nodes.push_back(krpc_node_to_dht_node(node));
+    }
+    return dht_nodes;
+}
+
+NodeId DhtClient::generate_node_id() {
+    NodeId id;
+    std::random_device rd;
+    std::mt19937 gen(rd());
+    std::uniform_int_distribution<> dis(0, 255);
+    
+    for (size_t i = 0; i < NODE_ID_SIZE; ++i) {
+        id[i] = dis(gen);
+    }
+    
+    return id;
+}
+
+NodeId DhtClient::xor_distance(const NodeId& a, const NodeId& b) {
+    NodeId result;
+    for (size_t i = 0; i < NODE_ID_SIZE; ++i) {
+        result[i] = a[i] ^ b[i];
+    }
+    return result;
+}
+
+bool DhtClient::is_closer(const NodeId& a, const NodeId& b, const NodeId& target) {
+    NodeId dist_a = xor_distance(a, target);
+    NodeId dist_b = xor_distance(b, target);
+    
+    return std::lexicographical_compare(dist_a.begin(), dist_a.end(),
+                                       dist_b.begin(), dist_b.end());
+}
+
+std::string DhtClient::generate_token(const Peer& peer) {
+    // Simple token generation (in real implementation, use proper cryptographic hash)
+    std::string data = peer.ip + ":" + std::to_string(peer.port);
+    std::hash<std::string> hasher;
+    size_t hash = hasher(data);
+    
+    // Convert hash to hex string
+    std::ostringstream oss;
+    oss << std::hex << hash;
+    std::string token = oss.str();
+    
+    // Store token for this peer with timestamp
+    {
+        std::lock_guard<std::mutex> lock(peer_tokens_mutex_);
+        peer_tokens_[peer] = PeerToken(token);
+    }
+    
+    return token;
+}
+
+bool DhtClient::verify_token(const Peer& peer, const std::string& token) {
+    std::lock_guard<std::mutex> lock(peer_tokens_mutex_);
+    auto it = peer_tokens_.find(peer);
+    if (it != peer_tokens_.end()) {
+        return it->second.token == token;
+    }
+    return false;
+}
+
+void DhtClient::cleanup_stale_nodes() {
+    std::lock_guard<std::mutex> routing_lock(routing_table_mutex_);
+    
+    auto now = std::chrono::steady_clock::now();
+    auto stale_threshold = std::chrono::minutes(15);
+    constexpr uint8_t MAX_FAIL_COUNT = 3;  // Remove after 3 consecutive failures
+    
+    size_t total_removed = 0;
+    
+    for (auto& bucket : routing_table_) {
+        auto old_size = bucket.size();
+        
+        bucket.erase(std::remove_if(bucket.begin(), bucket.end(),
+                                   [now, stale_threshold, MAX_FAIL_COUNT](const DhtNode& node) {
+                                       // Remove if too many failures
+                                       if (node.pinged() && node.fail_count >= MAX_FAIL_COUNT) {
+                                           LOG_DHT_DEBUG("Removing failed node " << node_id_to_hex(node.id) 
+                                                       << " (fail_count=" << static_cast<int>(node.fail_count) << ")");
+                                           return true;
+                                       }
+                                       
+                                       // Remove if never responded and too old
+                                       if (!node.pinged() && now - node.last_seen > stale_threshold) {
+                                           LOG_DHT_DEBUG("Removing unresponsive node " << node_id_to_hex(node.id) 
+                                                       << " (never responded, age > 15min)");
+                                           return true;
+                                       }
+                                       
+                                       return false;
+                                   }), bucket.end());
+        
+        total_removed += (old_size - bucket.size());
+    }
+    
+    if (total_removed > 0) {
+        LOG_DHT_DEBUG("Cleaned up " << total_removed << " stale/failed nodes from routing table");
+    }
+}
+
+void DhtClient::cleanup_stale_peer_tokens() {
+    std::lock_guard<std::mutex> lock(peer_tokens_mutex_);
+    
+    auto now = std::chrono::steady_clock::now();
+    auto stale_threshold = std::chrono::minutes(10);  // Tokens valid for 10 minutes (BEP 5 recommends tokens expire)
+    
+    size_t total_before = peer_tokens_.size();
+    
+    auto it = peer_tokens_.begin();
+    while (it != peer_tokens_.end()) {
+        if (now - it->second.created_at > stale_threshold) {
+            LOG_DHT_DEBUG("Removing stale token for peer " << it->first.ip << ":" << it->first.port);
+            it = peer_tokens_.erase(it);
+        } else {
+            ++it;
+        }
+    }
+    
+    size_t total_after = peer_tokens_.size();
+    
+    if (total_before > total_after) {
+        LOG_DHT_DEBUG("Cleaned up " << (total_before - total_after) << " stale peer tokens "
+                      << "(from " << total_before << " to " << total_after << ")");
+    }
+}
+
+void DhtClient::print_statistics() {
+    auto now = std::chrono::steady_clock::now();
+    
+    // Routing table statistics
+    size_t filled_buckets = 0;
+    size_t total_nodes = 0;
+    size_t max_bucket_size = 0;
+    size_t confirmed_nodes = 0;
+    size_t unpinged_nodes = 0;
+    size_t failed_nodes = 0;
+    
+    // Collect all nodes for best/worst analysis
+    std::vector<std::pair<DhtNode, int>> all_nodes;  // node + bucket index
+    
+    {
+        std::lock_guard<std::mutex> lock(routing_table_mutex_);
+        for (size_t bucket_idx = 0; bucket_idx < routing_table_.size(); ++bucket_idx) {
+            const auto& bucket = routing_table_[bucket_idx];
+            if (!bucket.empty()) {
+                filled_buckets++;
+                total_nodes += bucket.size();
+                max_bucket_size = (std::max)(max_bucket_size, bucket.size());
+                
+                for (const auto& node : bucket) {
+                    all_nodes.emplace_back(node, static_cast<int>(bucket_idx));
+                    
+                    if (node.confirmed()) {
+                        confirmed_nodes++;
+                    } else if (!node.pinged()) {
+                        unpinged_nodes++;
+                    } else if (node.fail_count > 0) {
+                        failed_nodes++;
+                    }
+                }
+            }
+        }
+    }
+    
+    // Pending searches statistics
+    size_t pending_searches = 0;
+    size_t total_search_nodes = 0;
+    size_t total_found_peers = 0;
+    size_t active_transactions = 0;
+    {
+        std::lock_guard<std::mutex> search_lock(pending_searches_mutex_);
+        pending_searches = pending_searches_.size();
+        active_transactions = transaction_to_search_.size();
+        for (const auto& [hash, search] : pending_searches_) {
+            total_search_nodes += search.search_nodes.size();
+            total_found_peers += search.found_peers.size();
+        }
+    }
+    
+    // Pending announces statistics
+    size_t pending_announces_count = 0;
+    {
+        std::lock_guard<std::mutex> announce_lock(pending_announces_mutex_);
+        pending_announces_count = pending_announces_.size();
+    }
+    
+    // Announced peers statistics
+    size_t announced_peers_total = 0;
+    size_t announced_peers_infohashes = 0;
+    {
+        std::lock_guard<std::mutex> peers_lock(announced_peers_mutex_);
+        announced_peers_infohashes = announced_peers_.size();
+        for (const auto& entry : announced_peers_) {
+            announced_peers_total += entry.second.size();
+        }
+    }
+    
+    // Ping verification statistics
+    size_t pending_pings = 0;
+    size_t nodes_being_replaced = 0;
+    {
+        std::lock_guard<std::mutex> ping_lock(pending_pings_mutex_);
+        pending_pings = pending_pings_.size();
+        nodes_being_replaced = nodes_being_replaced_.size();
+    }
+    
+    // Peer tokens statistics
+    size_t peer_tokens_count = 0;
+    {
+        std::lock_guard<std::mutex> tokens_lock(peer_tokens_mutex_);
+        peer_tokens_count = peer_tokens_.size();
+    }
+    
+    // Print main statistics
+    LOG_DHT_INFO("=== DHT GLOBAL STATISTICS ===");
+    LOG_DHT_INFO("[ROUTING TABLE]");
+    LOG_DHT_INFO("  Total nodes: " << total_nodes << " (confirmed: " << confirmed_nodes 
+                 << ", unpinged: " << unpinged_nodes << ", failed: " << failed_nodes << ")");
+    LOG_DHT_INFO("  Filled buckets: " << filled_buckets << "/" << routing_table_.size() 
+                 << ", Max bucket size: " << max_bucket_size << "/" << K_BUCKET_SIZE);
+    LOG_DHT_INFO("[ACTIVE OPERATIONS]");
+    LOG_DHT_INFO("  Pending searches: " << pending_searches 
+                 << " (nodes: " << total_search_nodes << ", found peers: " << total_found_peers << ")");
+    LOG_DHT_INFO("  Active transactions: " << active_transactions);
+    LOG_DHT_INFO("  Pending announces: " << pending_announces_count);
+    LOG_DHT_INFO("  Pending ping verifications: " << pending_pings 
+                 << " (nodes being replaced: " << nodes_being_replaced << ")");
+    LOG_DHT_INFO("[STORED DATA]");
+    LOG_DHT_INFO("  Announced peers: " << announced_peers_total 
+                 << " across " << announced_peers_infohashes << " infohashes");
+    LOG_DHT_INFO("  Peer tokens: " << peer_tokens_count);
+    
+    // Best/Worst nodes analysis
+    if (!all_nodes.empty()) {
+        // Sort by quality: confirmed first, then by RTT (lower is better)
+        std::sort(all_nodes.begin(), all_nodes.end(), 
+            [](const std::pair<DhtNode, int>& a, const std::pair<DhtNode, int>& b) {
+                // Confirmed nodes are better
+                if (a.first.confirmed() != b.first.confirmed()) {
+                    return a.first.confirmed();
+                }
+                // Lower fail_count is better
+                if (a.first.fail_count != b.first.fail_count) {
+                    return a.first.fail_count < b.first.fail_count;
+                }
+                // Lower RTT is better (0xffff = unknown, treat as worst)
+                return a.first.rtt < b.first.rtt;
+            });
+        
+        // Calculate RTT statistics (excluding unknown)
+        uint32_t rtt_sum = 0;
+        uint16_t rtt_min = 0xffff;
+        uint16_t rtt_max = 0;
+        size_t rtt_count = 0;
+        for (const auto& [node, bucket_idx] : all_nodes) {
+            if (node.rtt != 0xffff) {
+                rtt_sum += node.rtt;
+                rtt_min = (std::min)(rtt_min, node.rtt);
+                rtt_max = (std::max)(rtt_max, node.rtt);
+                rtt_count++;
+            }
+        }
+        
+        LOG_DHT_INFO("[RTT STATISTICS]");
+        if (rtt_count > 0) {
+            LOG_DHT_INFO("  Known RTT nodes: " << rtt_count << "/" << total_nodes);
+            LOG_DHT_INFO("  RTT min/avg/max: " << rtt_min << "ms / " 
+                         << (rtt_sum / rtt_count) << "ms / " << rtt_max << "ms");
+        } else {
+            LOG_DHT_INFO("  No RTT data available");
+        }
+        
+        // Show top 5 best nodes
+        LOG_DHT_INFO("[TOP 5 BEST NODES]");
+        size_t best_count = (std::min)(size_t(5), all_nodes.size());
+        for (size_t i = 0; i < best_count; ++i) {
+            const auto& [node, bucket_idx] = all_nodes[i];
+            auto age_seconds = std::chrono::duration_cast<std::chrono::seconds>(now - node.last_seen).count();
+            std::string status = node.confirmed() ? "confirmed" : 
+                                (node.pinged() ? "pinged" : "unpinged");
+            std::string rtt_str = (node.rtt == 0xffff) ? "N/A" : std::to_string(node.rtt) + "ms";
+            
+            LOG_DHT_INFO("  #" << (i + 1) << " " << node.peer.ip << ":" << node.peer.port 
+                         << " | bucket:" << bucket_idx << " rtt:" << rtt_str 
+                         << " fails:" << static_cast<int>(node.fail_count)
+                         << " " << status << " age:" << age_seconds << "s");
+        }
+        
+        // Show top 5 worst nodes
+        LOG_DHT_INFO("[TOP 5 WORST NODES]");
+        size_t worst_start = all_nodes.size() > 5 ? all_nodes.size() - 5 : 0;
+        for (size_t i = all_nodes.size(); i > worst_start; --i) {
+            const auto& [node, bucket_idx] = all_nodes[i - 1];
+            auto age_seconds = std::chrono::duration_cast<std::chrono::seconds>(now - node.last_seen).count();
+            std::string status = node.confirmed() ? "confirmed" : 
+                                (node.pinged() ? "pinged" : "unpinged");
+            std::string rtt_str = (node.rtt == 0xffff) ? "N/A" : std::to_string(node.rtt) + "ms";
+            
+            LOG_DHT_INFO("  #" << (all_nodes.size() - i + 1) << " " << node.peer.ip << ":" << node.peer.port 
+                         << " | bucket:" << bucket_idx << " rtt:" << rtt_str 
+                         << " fails:" << static_cast<int>(node.fail_count)
+                         << " " << status << " age:" << age_seconds << "s");
+        }
+    } else {
+        LOG_DHT_INFO("  No nodes in routing table");
+    }
+    LOG_DHT_INFO("=== END DHT STATISTICS ===");
+}
+
+void DhtClient::refresh_buckets() {
+    // Find random nodes in each bucket to refresh
+    std::lock_guard<std::mutex> lock(routing_table_mutex_);
+    
+    for (size_t i = 0; i < routing_table_.size(); ++i) {
+        if (routing_table_[i].empty()) {
+            // Generate a random node ID in this bucket's range
+            NodeId random_id = generate_node_id();
+            
+            // Set the appropriate bits to place it in bucket i
+            int byte_index = static_cast<int>(i / 8);
+            int bit_index = static_cast<int>(i % 8);
+            
+            if (byte_index < NODE_ID_SIZE) {
+                // Clear the target bit and higher bits
+                for (int j = byte_index; j < NODE_ID_SIZE; ++j) {
+                    random_id[j] = node_id_[j];
+                }
+                
+                // Set the target bit
+                random_id[byte_index] |= (1 << (7 - bit_index));
+                
+                // Find nodes to query
+                auto closest_nodes = find_closest_nodes_unlocked(random_id, ALPHA);
+                for (const auto& node : closest_nodes) {
+                    send_krpc_find_node(node.peer, random_id);
+                }
+            }
+        }
+    }
+}
+
+void DhtClient::cleanup_stale_announces() {
+    std::lock_guard<std::mutex> lock(pending_announces_mutex_);
+    
+    auto now = std::chrono::steady_clock::now();
+    auto stale_threshold = std::chrono::minutes(5);  // Remove announces older than 5 minutes
+    
+    auto it = pending_announces_.begin();
+    while (it != pending_announces_.end()) {
+        if (now - it->second.created_at > stale_threshold) {
+            LOG_DHT_DEBUG("Removing stale pending announce for transaction " << it->first);
+            it = pending_announces_.erase(it);
+        } else {
+            ++it;
+        }
+    }
+}
+
+void DhtClient::cleanup_stale_searches() {
+    std::lock_guard<std::mutex> lock(pending_searches_mutex_);
+    
+    auto now = std::chrono::steady_clock::now();
+    auto stale_threshold = std::chrono::minutes(5);  // Remove searches older than 5 minutes
+    
+    // Clean up stale searches (by info_hash)
+    auto search_it = pending_searches_.begin();
+    while (search_it != pending_searches_.end()) {
+        if (now - search_it->second.created_at > stale_threshold) {
+            LOG_DHT_DEBUG("Removing stale pending search for info_hash " << search_it->first);
+            search_it = pending_searches_.erase(search_it);
+        } else {
+            ++search_it;
+        }
+    }
+    
+    // Clean up stale transaction mappings (remove ones that point to non-existent searches)
+    auto trans_it = transaction_to_search_.begin();
+    while (trans_it != transaction_to_search_.end()) {
+        if (pending_searches_.find(trans_it->second.info_hash_hex) == pending_searches_.end()) {
+            LOG_DHT_DEBUG("Removing stale transaction mapping " << trans_it->first << " -> " << trans_it->second.info_hash_hex);
+            trans_it = transaction_to_search_.erase(trans_it);
+        } else {
+            ++trans_it;
+        }
+    }
+}
+
+void DhtClient::cleanup_search_node_states() {
+    std::lock_guard<std::mutex> lock(pending_searches_mutex_);
+    constexpr size_t MAX_NODE_STATES = 200;  // Twice bigger than MAX_SEARCH_NODES
+
+    for (auto& [hash_key, search] : pending_searches_) {
+        if (search.is_finished) {
+            continue;
+        }
+
+        // Skip cleanup if node_states is within acceptable limits
+        if (search.node_states.size() <= MAX_NODE_STATES) {
+            continue;
+        }
+
+        // Erase while iterating using iterator-based loop
+        size_t removed = 0;
+        auto it = search.node_states.begin();
+        while (it != search.node_states.end()) {
+            // В cleanup_search_node_states:
+            if ((it->second & SearchNodeFlags::ABANDONED) && 
+            ((it->second & (SearchNodeFlags::TIMED_OUT | SearchNodeFlags::RESPONDED)) ||
+            !(it->second & SearchNodeFlags::QUERIED))) {  // Never queried = safe to remove immediately
+                it = search.node_states.erase(it);
+                removed++;
+            } else {
+                ++it;
+            }
+        }
+        
+        if (removed > 0) {
+            LOG_DHT_DEBUG("Cleaned up " << removed << " abandoned nodes for search " << hash_key 
+                          << " (remaining: " << search.node_states.size() << ")");
+        }
+    }
+}
+
+void DhtClient::cleanup_timed_out_search_requests() {
+    std::vector<DeferredCallbacks> all_deferred;
+    
+    {
+        std::lock_guard<std::mutex> lock(pending_searches_mutex_);
+
+        if (pending_searches_.empty()) {
+            return;
+        }
+        
+        auto now = std::chrono::steady_clock::now();
+        // - Short timeout (2s): Free up the slot by increasing branch_factor, but keep waiting for late response
+        // - Full timeout (15s): Mark node as failed and remove the transaction
+        auto short_timeout_threshold = std::chrono::seconds(2);
+        auto full_timeout_threshold = std::chrono::seconds(15);
+        
+        // Collect transactions that need short timeout or full timeout
+        std::vector<std::string> short_timeout_transactions;
+        std::vector<std::string> full_timeout_transactions;
+        
+        for (const auto& [transaction_id, trans_info] : transaction_to_search_) {
+            auto elapsed = now - trans_info.sent_at;
+            
+            if (elapsed > full_timeout_threshold) {
+                full_timeout_transactions.push_back(transaction_id);
+            } else if (elapsed > short_timeout_threshold) {
+                // Check if this node already has short timeout
+                auto search_it = pending_searches_.find(trans_info.info_hash_hex);
+                if (search_it != pending_searches_.end()) {
+                    auto& search = search_it->second;
+                    // Only process if not already marked with short timeout
+                    auto state_it = search.node_states.find(trans_info.queried_node_id);
+                    if (state_it == search.node_states.end() || !(state_it->second & SearchNodeFlags::SHORT_TIMEOUT)) {
+                        short_timeout_transactions.push_back(transaction_id);
+                    }
+                }
+            }
+        }
+        
+        // Group by search to batch process and call add_search_requests once per search
+        std::unordered_set<std::string> affected_searches;
+        
+        // Process short timeouts first - these nodes are slow but we still wait for a response
+        for (const auto& transaction_id : short_timeout_transactions) {
+            auto trans_it = transaction_to_search_.find(transaction_id);
+            if (trans_it == transaction_to_search_.end()) {
+                continue;
+            }
+            
+            const auto& trans_info = trans_it->second;
+            auto search_it = pending_searches_.find(trans_info.info_hash_hex);
+            
+            if (search_it != pending_searches_.end()) {
+                auto& search = search_it->second;
+                
+                if (!search.is_finished) {
+                    // Check if this node was abandoned during truncation
+                    auto state_it = search.node_states.find(trans_info.queried_node_id);
+                    if (state_it != search.node_states.end() && 
+                        (state_it->second & SearchNodeFlags::ABANDONED)) {
+                        // Node was abandoned, skip short timeout processing
+                        continue;
+                    }
+                    
+                    // Mark node with short timeout (add flag, preserving existing flags)
+                    search.node_states[trans_info.queried_node_id] |= SearchNodeFlags::SHORT_TIMEOUT;
+                    
+                    // Increase branch factor to allow another request (opening up a slot)
+                    search.branch_factor++;
+                    
+                    LOG_DHT_DEBUG("Short timeout for node " << node_id_to_hex(trans_info.queried_node_id) 
+                                  << " in search " << trans_info.info_hash_hex 
+                                  << " - increased branch_factor to " << search.branch_factor
+                                  << " (still waiting for late response)");
+                    
+                    affected_searches.insert(trans_info.info_hash_hex);
+                }
+            }
+            // Note: We DON'T remove the transaction - we're still waiting for a possible late response
+        }
+        
+        // Process full timeouts - these nodes have completely failed
+        for (const auto& transaction_id : full_timeout_transactions) {
+            auto trans_it = transaction_to_search_.find(transaction_id);
+            if (trans_it == transaction_to_search_.end()) {
+                continue;
+            }
+            
+            const auto& trans_info = trans_it->second;
+            auto search_it = pending_searches_.find(trans_info.info_hash_hex);
+            
+            if (search_it != pending_searches_.end()) {
+                auto& search = search_it->second;
+                
+                if (!search.is_finished) {
+                    // Get current flags for this node
+                    uint8_t& flags = search.node_states[trans_info.queried_node_id];
+                    
+                    // Check if this node was abandoned during truncation
+                    if (flags & SearchNodeFlags::ABANDONED) {
+                        // Node was abandoned, invoke_count already decremented
+                        // Mark as timed out so cleanup_search_node_states can remove it from node_states
+                        flags |= SearchNodeFlags::TIMED_OUT;
+                        transaction_to_search_.erase(trans_it);
+                        continue;
+                    }
+                    
+                    bool had_short_timeout = flags & SearchNodeFlags::SHORT_TIMEOUT;
+                    
+                    // Always decrement invoke_count on full timeout (node was still in-flight)
+                    if (search.invoke_count > 0) {
+                        search.invoke_count--;
+                    }
+                    
+                    if (had_short_timeout) {
+                        // Restore branch factor since node fully timed out
+                        if (search.branch_factor > static_cast<int>(ALPHA)) {
+                            search.branch_factor--;
+                        }
+                        
+                        LOG_DHT_DEBUG("Full timeout for node " << node_id_to_hex(trans_info.queried_node_id) 
+                                      << " in search " << trans_info.info_hash_hex 
+                                      << " (had short timeout) - restored branch_factor to " << search.branch_factor
+                                      << ", invoke_count now: " << search.invoke_count);
+                    } else {
+                        LOG_DHT_DEBUG("Full timeout for node " << node_id_to_hex(trans_info.queried_node_id) 
+                                      << " in search " << trans_info.info_hash_hex 
+                                      << " - invoke_count now: " << search.invoke_count);
+                    }
+                    
+                    // Mark the node as timed out (add flag, preserving history)
+                    flags |= SearchNodeFlags::TIMED_OUT;
+                    
+                    // Mark the node as failed in routing table (BEP 5 compliance)
+                    {
+                        std::lock_guard<std::mutex> rt_lock(routing_table_mutex_);
+                        int bucket_index = get_bucket_index(trans_info.queried_node_id);
+                        auto& bucket = routing_table_[bucket_index];
+                        auto node_it = std::find_if(bucket.begin(), bucket.end(),
+                            [&trans_info](const DhtNode& n) { return n.id == trans_info.queried_node_id; });
+                        if (node_it != bucket.end()) {
+                            node_it->mark_failed();
+                            LOG_DHT_DEBUG("Marked node " << node_id_to_hex(trans_info.queried_node_id) 
+                                          << " as failed in routing table (fail_count=" 
+                                          << static_cast<int>(node_it->fail_count) << ")");
+                        }
+                    }
+                    
+                    affected_searches.insert(trans_info.info_hash_hex);
+                }
+            }
+            
+            // Remove the fully timed out transaction
+            transaction_to_search_.erase(trans_it);
+        }
+        
+        if (!short_timeout_transactions.empty() || !full_timeout_transactions.empty()) {
+            LOG_DHT_DEBUG("Timeout handling: " << short_timeout_transactions.size() << " short timeouts, "
+                          << full_timeout_transactions.size() << " full timeouts");
+        }
+        
+        // Continue searches that had timeout events
+        for (const auto& hash_key : affected_searches) {
+            auto search_it = pending_searches_.find(hash_key);
+            if (search_it != pending_searches_.end() && !search_it->second.is_finished) {
+                LOG_DHT_DEBUG("Continuing search " << hash_key << " after timeout handling");
+                DeferredCallbacks deferred;
+                add_search_requests(search_it->second, deferred);
+                if (deferred.should_invoke) {
+                    all_deferred.push_back(std::move(deferred));
+                }
+            }
+        }
+        
+        // Clean up finished searches
+        for (const auto& hash_key : affected_searches) {
+            auto search_it = pending_searches_.find(hash_key);
+            if (search_it != pending_searches_.end() && search_it->second.is_finished) {
+                LOG_DHT_DEBUG("Removing finished search " << hash_key << " after timeout handling");
+                pending_searches_.erase(search_it);
+            }
+        }
+    }
+    
+    // Invoke all deferred callbacks outside the lock to avoid deadlock
+    for (auto& deferred : all_deferred) {
+        deferred.invoke();
+    }
+}
+
+void DhtClient::handle_get_peers_response_for_announce(const std::string& transaction_id, const Peer& responder, const std::string& token) {
+    std::lock_guard<std::mutex> lock(pending_announces_mutex_);
+    
+    auto it = pending_announces_.find(transaction_id);
+    if (it != pending_announces_.end()) {
+        const auto& pending_announce = it->second;
+        LOG_DHT_DEBUG("Found pending announce for transaction " << transaction_id 
+                      << " - sending announce_peer for info_hash " << node_id_to_hex(pending_announce.info_hash) 
+                      << " to " << responder.ip << ":" << responder.port);
+        
+        // Send announce_peer with the received token
+        send_krpc_announce_peer(responder, pending_announce.info_hash, pending_announce.port, token);
+        
+        // Remove the pending announce since we've handled it
+        pending_announces_.erase(it);
+    }
+}
+
+void DhtClient::handle_get_peers_empty_response(const std::string& transaction_id, const Peer& responder) {
+    DeferredCallbacks deferred;
+    {
+        std::lock_guard<std::mutex> lock(pending_searches_mutex_);
+        auto trans_it = transaction_to_search_.find(transaction_id);
+        if (trans_it != transaction_to_search_.end()) {
+            const auto& trans_info = trans_it->second;
+            auto search_it = pending_searches_.find(trans_info.info_hash_hex);
+            if (search_it != pending_searches_.end()) {
+                auto& pending_search = search_it->second;
+
+                // Check if this node was abandoned during truncation
+                auto state_it = pending_search.node_states.find(trans_info.queried_node_id);
+                if (state_it != pending_search.node_states.end() && 
+                    (state_it->second & SearchNodeFlags::ABANDONED)) {
+                    // Mark as responded so cleanup_search_node_states can remove it
+                    state_it->second |= SearchNodeFlags::RESPONDED;
+                    LOG_DHT_DEBUG("Ignoring empty response from abandoned node " 
+                                  << node_id_to_hex(trans_info.queried_node_id));
+                    return;
+                }
+
+                uint8_t& flags = pending_search.node_states[trans_info.queried_node_id];
+
+                if (flags & SearchNodeFlags::RESPONDED) {
+                    LOG_DHT_DEBUG("Ignoring duplicate response from node " << node_id_to_hex(trans_info.queried_node_id));
+                    return;
+                }
+                
+                // Decrement invoke count
+                if (pending_search.invoke_count > 0) {
+                    pending_search.invoke_count--;
+                }
+                
+                // Restore branch_factor if had short timeout
+                if (flags & SearchNodeFlags::SHORT_TIMEOUT) {
+                    if (pending_search.branch_factor > static_cast<int>(ALPHA)) {
+                        pending_search.branch_factor--;
+                    }
+                }
+                
+                // Mark as responded
+                flags |= SearchNodeFlags::RESPONDED;
+                
+                LOG_DHT_DEBUG("Empty get_peers response from " << responder.ip << ":" << responder.port
+                              << " for info_hash " << trans_info.info_hash_hex
+                              << " (invoke_count now: " << pending_search.invoke_count << ")");
+                
+                // Continue search
+                add_search_requests(pending_search, deferred);
+            }
+        }
+    }
+    
+    deferred.invoke();
+}
+
+void DhtClient::handle_get_peers_response_for_search(const std::string& transaction_id, const Peer& responder, const std::vector<Peer>& peers) {
+    DeferredCallbacks deferred_immediate;   // For new peers callbacks
+    DeferredCallbacks deferred_completion;  // For search completion callbacks
+
+    {
+        std::lock_guard<std::mutex> lock(pending_searches_mutex_);
+        auto trans_it = transaction_to_search_.find(transaction_id);
+        if (trans_it != transaction_to_search_.end()) {
+            const auto& trans_info = trans_it->second;
+            auto search_it = pending_searches_.find(trans_info.info_hash_hex);
+            if (search_it != pending_searches_.end()) {
+                auto& pending_search = search_it->second;
+
+                // Check if this node was abandoned during truncation
+                auto state_it = pending_search.node_states.find(trans_info.queried_node_id);
+                if (state_it != pending_search.node_states.end() && 
+                    (state_it->second & SearchNodeFlags::ABANDONED)) {
+                    // Mark as responded so cleanup_search_node_states can remove it
+                    state_it->second |= SearchNodeFlags::RESPONDED;
+                    LOG_DHT_DEBUG("Ignoring response from abandoned node " 
+                                << node_id_to_hex(trans_info.queried_node_id)
+                                << " - invoke_count already decremented during truncation");
+                    return;
+                }
+
+                // Get flags for this node and mark as responded
+                uint8_t& flags = pending_search.node_states[trans_info.queried_node_id];
+
+                // Check if already responded (duplicate response)
+                if (flags & SearchNodeFlags::RESPONDED) {
+                    LOG_DHT_DEBUG("Ignoring duplicate response from node " 
+                                << node_id_to_hex(trans_info.queried_node_id));
+                    return;
+                }
+                
+                // Decrement invoke count since we received a response
+                if (pending_search.invoke_count > 0) {
+                    pending_search.invoke_count--;
+                }
+                
+                // If this node had short timeout, restore the branch factor (late response arrived)
+                if (flags & SearchNodeFlags::SHORT_TIMEOUT) {
+                    if (pending_search.branch_factor > static_cast<int>(ALPHA)) {
+                        pending_search.branch_factor--;
+                    }
+                    LOG_DHT_DEBUG("Late response from node " << node_id_to_hex(trans_info.queried_node_id)
+                                << " (had short timeout) - restored branch_factor to " << pending_search.branch_factor);
+                }
+                
+                // Mark as responded (add flag, preserving history including SHORT_TIMEOUT)
+                flags |= SearchNodeFlags::RESPONDED;
+                
+                LOG_DHT_DEBUG("Found pending search for KRPC transaction " << transaction_id 
+                            << " - received " << peers.size() << " peers for info_hash " << trans_info.info_hash_hex 
+                            << " from " << responder.ip << ":" << responder.port
+                            << " (invoke_count now: " << pending_search.invoke_count << ")");
+
+                // Accumulate peers (with deduplication) - continue search like reference implementation
+                if (!peers.empty()) {
+                    // Collect only new (non-duplicate) peers for immediate callback
+                    std::vector<Peer> new_peers;
+                    new_peers.reserve(peers.size());
+                    
+                    for (const auto& peer : peers) {
+                        // Check if peer already exists in found_peers
+                        auto it = std::find_if(pending_search.found_peers.begin(), 
+                                            pending_search.found_peers.end(),
+                                            [&peer](const Peer& p) { 
+                                                return p.ip == peer.ip && p.port == peer.port; 
+                                            });
+                        if (it == pending_search.found_peers.end()) {
+                            pending_search.found_peers.push_back(peer);
+                            new_peers.push_back(peer);
+                            LOG_DHT_DEBUG("  [new] found peer for hash(" << trans_info.info_hash_hex << ") = " << peer.ip << ":" << peer.port);
+                        }
+                    }
+                    
+                    // Collect immediate callbacks for new peers
+                    if (!new_peers.empty()) {
+                        LOG_DHT_DEBUG("Invoking " << pending_search.callbacks.size() << " callbacks with " 
+                                    << new_peers.size() << " new peers for info_hash " << trans_info.info_hash_hex);
+                        deferred_immediate.should_invoke = true;
+                        deferred_immediate.peers = std::move(new_peers);
+                        deferred_immediate.info_hash = pending_search.info_hash;
+                        deferred_immediate.callbacks = pending_search.callbacks;
+                    }
+                    
+                    LOG_DHT_DEBUG("Accumulated " << pending_search.found_peers.size() << " total peers for info_hash " << trans_info.info_hash_hex);
+                }
+                
+                // Continue search - let add_search_requests determine when to finish
+                add_search_requests(pending_search, deferred_completion);
+            }
+            
+            // DON'T remove the transaction mapping here - it will be removed at the end of handle_krpc_response
+            // This ensures all response data is fully processed before cleanup
+        }
+    }
+
+    // Invoke all callbacks outside the lock to avoid deadlock
+    deferred_immediate.invoke();
+    deferred_completion.invoke();
+}
+
+
+void DhtClient::handle_get_peers_response_with_nodes(const std::string& transaction_id, const Peer& responder, const std::vector<KrpcNode>& nodes) {
+    // This function is called when get_peers returns nodes instead of peers
+    // Add the new nodes to search_nodes and continue the search
+    
+    DeferredCallbacks deferred;
+    
+    {
+        std::lock_guard<std::mutex> lock(pending_searches_mutex_);
+        
+        auto trans_it = transaction_to_search_.find(transaction_id);
+        if (trans_it != transaction_to_search_.end()) {
+            const auto& trans_info = trans_it->second;
+            auto search_it = pending_searches_.find(trans_info.info_hash_hex);
+            if (search_it != pending_searches_.end()) {
+                auto& pending_search = search_it->second;
+
+                // Check if this node was abandoned during truncation
+                auto state_it = pending_search.node_states.find(trans_info.queried_node_id);
+                if (state_it != pending_search.node_states.end() && 
+                    (state_it->second & SearchNodeFlags::ABANDONED)) {
+                    // Mark as responded so cleanup_search_node_states can remove it
+                    state_it->second |= SearchNodeFlags::RESPONDED;
+                    LOG_DHT_DEBUG("Ignoring response from abandoned node " 
+                                  << node_id_to_hex(trans_info.queried_node_id)
+                                  << " - invoke_count already decremented during truncation");
+                    return;
+                }
+
+                // Get flags for this node and mark as responded
+                uint8_t& flags = pending_search.node_states[trans_info.queried_node_id];
+
+                // Check if already responded (duplicate response)
+                if (flags & SearchNodeFlags::RESPONDED) {
+                    LOG_DHT_DEBUG("Ignoring duplicate response from node " 
+                                << node_id_to_hex(trans_info.queried_node_id));
+                    return;
+                }
+                
+                // Decrement invoke count since we received a response
+                if (pending_search.invoke_count > 0) {
+                    pending_search.invoke_count--;
+                }
+                
+                // If this node had short timeout, restore the branch factor (late response arrived)
+                if (flags & SearchNodeFlags::SHORT_TIMEOUT) {
+                    if (pending_search.branch_factor > static_cast<int>(ALPHA)) {
+                        pending_search.branch_factor--;
+                    }
+                    LOG_DHT_DEBUG("Late response from node " << node_id_to_hex(trans_info.queried_node_id)
+                                  << " (had short timeout) - restored branch_factor to " << pending_search.branch_factor);
+                }
+                
+                // Mark as responded (add flag, preserving history including SHORT_TIMEOUT)
+                flags |= SearchNodeFlags::RESPONDED;
+                
+                LOG_DHT_DEBUG("Processing get_peers response with " << nodes.size() 
+                              << " nodes for info_hash " << trans_info.info_hash_hex << " from " << responder.ip << ":" << responder.port
+                              << " (invoke_count now: " << pending_search.invoke_count << ")");
+                
+                // Add new nodes to search_nodes (sorted by distance)
+                size_t nodes_added = 0;
+                for (const auto& node : nodes) {
+                    DhtNode dht_node = krpc_node_to_dht_node(node);
+                    size_t old_size = pending_search.search_nodes.size();
+                    add_node_to_search(pending_search, dht_node);
+                    if (pending_search.search_nodes.size() > old_size) {
+                        nodes_added++;
+                    }
+                }
+                
+                LOG_DHT_DEBUG("Added " << nodes_added << " new nodes to search_nodes (total: " << pending_search.search_nodes.size() << ")");
+                
+                // Continue search with new nodes
+                add_search_requests(pending_search, deferred);
+            }
+            
+            // DON'T remove the transaction mapping here - it will be removed at the end of handle_krpc_response
+            // This ensures all response data is fully processed before cleanup
+        }
+    }
+    
+    // Invoke callbacks outside the lock to avoid deadlock
+    deferred.invoke();
+}
+
+
+void DhtClient::add_node_to_search(PendingSearch& search, const DhtNode& node) {
+    // Check if node already exists in search (node is "known" if it's in node_states map)
+    if (search.node_states.find(node.id) != search.node_states.end()) {
+        LOG_DHT_DEBUG("Node " << node_id_to_hex(node.id) << " already known for search - skipping");
+        return;
+    }
+    
+    // Find insertion point to maintain sorted order (closest first)
+    auto insert_pos = std::lower_bound(search.search_nodes.begin(), search.search_nodes.end(), node,
+                                       [&search, this](const DhtNode& a, const DhtNode& b) {
+                                           return is_closer(a.id, b.id, search.info_hash);
+                                       });
+    
+    search.search_nodes.insert(insert_pos, node);
+    // Mark node as known (add to map with no flags set - will get QUERIED flag when query is sent)
+    search.node_states[node.id] = 0;
+    
+    // Limit search_nodes size to avoid unbounded growth
+    constexpr size_t MAX_SEARCH_NODES = 100;
+    if (search.search_nodes.size() > MAX_SEARCH_NODES) {
+        // Before truncating, clean up counters for in-flight queries being discarded
+        for (size_t i = MAX_SEARCH_NODES; i < search.search_nodes.size(); ++i) {
+            const auto& discarded_node = search.search_nodes[i];
+            auto state_it = search.node_states.find(discarded_node.id);
+            if (state_it != search.node_states.end()) {
+                uint8_t flags = state_it->second;
+                // If queried but not responded/failed, it's in-flight
+                if ((flags & SearchNodeFlags::QUERIED) && 
+                    !(flags & (SearchNodeFlags::RESPONDED | SearchNodeFlags::TIMED_OUT))) {
+                    // Decrement invoke_count since this request is being abandoned
+                    if (search.invoke_count > 0) {
+                        search.invoke_count--;
+                        LOG_DHT_DEBUG("Decrementing invoke_count for abandoned node " 
+                                      << node_id_to_hex(discarded_node.id) 
+                                      << " (now: " << search.invoke_count << ")");
+                    }
+                    // If it had short timeout, also restore branch factor
+                    if (flags & SearchNodeFlags::SHORT_TIMEOUT) {
+                        if (search.branch_factor > static_cast<int>(ALPHA)) {
+                            search.branch_factor--;
+                            LOG_DHT_DEBUG("Decrementing branch_factor for abandoned node with short_timeout " 
+                                          << node_id_to_hex(discarded_node.id) 
+                                          << " (now: " << search.branch_factor << ")");
+                        }
+                    }
+                }
+                // Mark as ABANDONED instead of removing - prevents double invoke_count decrement
+                // when late response arrives (response handlers check this flag)
+                state_it->second |= SearchNodeFlags::ABANDONED;
+            }
+        }
+        search.search_nodes.resize(MAX_SEARCH_NODES);
+    }
+}
+
+bool DhtClient::add_search_requests(PendingSearch& search, DeferredCallbacks& deferred) {
+    // Returns true if search is done (completed or should be finished)
+    
+    if (search.is_finished) {
+        return true;
+    }
+    
+    std::string hash_key = node_id_to_hex(search.info_hash);
+    
+    LOG_DHT_DEBUG("Adding search requests for info_hash " << hash_key);
+    
+    const int k = static_cast<int>(K_BUCKET_SIZE);  // Target number of results
+    int loop_index = -1;
+    int results_found = 0;       // Nodes that have responded
+    int queries_in_flight = 0;   // Requests currently in flight
+    int timed_out_count = 0;     // Nodes that timed out
+    int queries_sent = 0;        // Queries sent this round
+    
+    // Iterate through search_nodes (sorted by distance, closest first)
+    // Important: We must continue iterating to count results even when we can't send more requests
+    for (auto& node : search.search_nodes) {
+        loop_index++;
+
+        // Stop if we have enough completed results
+        if (results_found >= k) {
+            break;
+        }
+
+        // Get flags for this node (0 if not in map, meaning just "known")
+        auto state_it = search.node_states.find(node.id);
+        uint8_t flags = (state_it != search.node_states.end()) ? state_it->second : 0;
+
+        // Usually it doesn't happen, but if it does, we skip it
+        if (flags & SearchNodeFlags::ABANDONED) {
+            continue;
+        }
+        
+        // Check if this node has already responded (counts toward results)
+        if (flags & SearchNodeFlags::RESPONDED) {
+            results_found++;
+            continue;
+        }
+        
+        // Skip nodes that have timed out (don't count as results or in-flight)
+        if (flags & SearchNodeFlags::TIMED_OUT) {
+            timed_out_count++;
+            continue;
+        }
+        
+        // Check if this node was already queried
+        if (flags & SearchNodeFlags::QUERIED) {
+            // Only count as in-flight if not responded yet
+            // (TIMED_OUT already handled above, RESPONDED handled above too)
+            // This case handles nodes that are QUERIED but still waiting for response
+            queries_in_flight++;
+            continue;
+        }
+        
+        // Check if we have capacity to send more requests
+        // Important: use 'continue' not 'break' to keep counting results
+        // Use adaptive branch_factor (increases on short timeout, restores on response/full timeout)
+        if (search.invoke_count >= search.branch_factor) {
+            continue;
+        }
+        
+        // Send query to this node
+        std::string transaction_id = KrpcProtocol::generate_transaction_id();
+        transaction_to_search_[transaction_id] = SearchTransaction(hash_key, node.id);
+        search.node_states[node.id] |= SearchNodeFlags::QUERIED;
+        search.invoke_count++;
+        
+        LOG_DHT_DEBUG("Querying node " << node_id_to_hex(node.id) << " at " << node.peer.ip << ":" << node.peer.port);
+        
+        auto message = KrpcProtocol::create_get_peers_query(transaction_id, node_id_, search.info_hash);
+        send_krpc_message(message, node.peer);
+        
+        queries_sent++;
+    }
+    
+    LOG_DHT_DEBUG("Search [" << hash_key << "] progress [ms: " << std::chrono::duration_cast<std::chrono::milliseconds>(std::chrono::steady_clock::now() - search.created_at).count() << "]:");
+    LOG_DHT_DEBUG(" * search_nodes: " << search.search_nodes.size());
+    LOG_DHT_DEBUG(" * queries_sent: " << queries_sent);
+    LOG_DHT_DEBUG(" * invoke_count: " << search.invoke_count);
+    LOG_DHT_DEBUG(" * branch_factor: " << search.branch_factor);
+    LOG_DHT_DEBUG(" * results_found: " << results_found);
+    LOG_DHT_DEBUG(" * queries_in_flight: " << queries_in_flight);
+    LOG_DHT_DEBUG(" * timed_out: " << timed_out_count);
+    LOG_DHT_DEBUG(" * peers_found: " << search.found_peers.size());
+    LOG_DHT_DEBUG(" * callbacks: " << search.callbacks.size());
+    LOG_DHT_DEBUG(" * loop_index: " << loop_index);
+    LOG_DHT_DEBUG(" * node_states: " << search.node_states.size());
+    
+    if ((results_found >= k && queries_in_flight == 0) || search.invoke_count == 0) {
+        auto duration_ms = std::chrono::duration_cast<std::chrono::milliseconds>(
+            std::chrono::steady_clock::now() - search.created_at
+        ).count();
+        
+        // Count final stats for completion log
+        int queried_total = 0, responded_total = 0, timed_out_total = 0, short_timeout_total = 0, abandoned_total = 0;
+        for (const auto& [id, f] : search.node_states) {
+            if (f & SearchNodeFlags::QUERIED) queried_total++;
+            if (f & SearchNodeFlags::RESPONDED) responded_total++;
+            if (f & SearchNodeFlags::TIMED_OUT) timed_out_total++;
+            if (f & SearchNodeFlags::SHORT_TIMEOUT) short_timeout_total++;
+            if (f & SearchNodeFlags::ABANDONED) abandoned_total++;
+        }
+        
+        LOG_DHT_INFO("=== Search Completed for info_hash " << hash_key << " ===");
+        LOG_DHT_INFO("  Duration: " << duration_ms << "ms");
+        LOG_DHT_INFO("  Total nodes queried: " << queried_total);
+        LOG_DHT_INFO("  Total nodes responded: " << responded_total);
+        LOG_DHT_INFO("  Total nodes timed out: " << timed_out_total);
+        LOG_DHT_INFO("  Nodes with short timeout: " << short_timeout_total);
+        LOG_DHT_INFO("  Nodes abandoned (truncation): " << abandoned_total);
+        LOG_DHT_INFO("  Final branch_factor: " << search.branch_factor << " (initial: " << ALPHA << ")");
+        LOG_DHT_INFO("  Total peers found: " << search.found_peers.size());
+        LOG_DHT_INFO("  Callbacks to invoke: " << search.callbacks.size());
+        
+        // Collect callbacks for deferred invocation (avoid deadlock - don't call user callbacks while holding mutex)
+        deferred.should_invoke = true;
+        deferred.callbacks = search.callbacks;
+        deferred.peers = search.found_peers;
+        deferred.info_hash = search.info_hash;
+        
+        search.is_finished = true;
+        return true;
+    }
+    
+    return false;
+}
+
+// Peer announcement storage management
+void DhtClient::store_announced_peer(const InfoHash& info_hash, const Peer& peer) {
+    std::lock_guard<std::mutex> lock(announced_peers_mutex_);
+    
+    std::string hash_key = node_id_to_hex(info_hash);
+    auto& peers = announced_peers_[hash_key];
+    
+    // Check if peer already exists
+    auto it = std::find_if(peers.begin(), peers.end(),
+                          [&peer](const AnnouncedPeer& announced) {
+                              return announced.peer.ip == peer.ip && announced.peer.port == peer.port;
+                          });
+    
+    if (it != peers.end()) {
+        // Update existing peer's timestamp
+        it->announced_at = std::chrono::steady_clock::now();
+        LOG_DHT_DEBUG("Updated existing announced peer " << peer.ip << ":" << peer.port 
+                      << " for info_hash " << hash_key);
+    } else {
+        // Add new peer
+        peers.emplace_back(peer);
+        LOG_DHT_DEBUG("Stored new announced peer " << peer.ip << ":" << peer.port 
+                      << " for info_hash " << hash_key << " (total: " << peers.size() << ")");
+    }
+}
+
+std::vector<Peer> DhtClient::get_announced_peers(const InfoHash& info_hash) {
+    std::lock_guard<std::mutex> lock(announced_peers_mutex_);
+    
+    std::string hash_key = node_id_to_hex(info_hash);
+    auto it = announced_peers_.find(hash_key);
+    
+    std::vector<Peer> peers;
+    if (it != announced_peers_.end()) {
+        peers.reserve(it->second.size());
+        for (const auto& announced : it->second) {
+            peers.push_back(announced.peer);
+        }
+        LOG_DHT_DEBUG("Retrieved " << peers.size() << " announced peers for info_hash " << hash_key);
+    } else {
+        LOG_DHT_DEBUG("No announced peers found for info_hash " << hash_key);
+    }
+    
+    return peers;
+}
+
+void DhtClient::cleanup_stale_announced_peers() {
+    std::lock_guard<std::mutex> lock(announced_peers_mutex_);
+    
+    auto now = std::chrono::steady_clock::now();
+    auto stale_threshold = std::chrono::minutes(30);  // BEP 5 standard: 30 minutes
+    
+    size_t total_before = 0;
+    size_t total_after = 0;
+    
+    for (auto it = announced_peers_.begin(); it != announced_peers_.end(); ) {
+        auto& peers = it->second;
+        total_before += peers.size();
+        
+        // Remove stale peers
+        peers.erase(std::remove_if(peers.begin(), peers.end(),
+                                   [now, stale_threshold](const AnnouncedPeer& announced) {
+                                       return now - announced.announced_at > stale_threshold;
+                                   }), peers.end());
+        
+        total_after += peers.size();
+        
+        // Remove empty info_hash entries
+        if (peers.empty()) {
+            LOG_DHT_DEBUG("Removing empty announced peers entry for info_hash " << it->first);
+            it = announced_peers_.erase(it);
+        } else {
+            ++it;
+        }
+    }
+    
+    if (total_before > total_after) {
+        LOG_DHT_DEBUG("Cleaned up " << (total_before - total_after) << " stale announced peers "
+                      << "(from " << total_before << " to " << total_after << ")");
+    }
+}
+
+// Ping-before-replace eviction implementation
+void DhtClient::initiate_ping_verification(const DhtNode& candidate_node, const DhtNode& old_node, int bucket_index) {
+    // NOTE: pending_pings_mutex_ is already held by caller (add_node)
+    
+    std::string ping_transaction_id = KrpcProtocol::generate_transaction_id();
+    
+    LOG_DHT_DEBUG("Initiating ping verification: pinging OLD node " << node_id_to_hex(old_node.id) 
+                  << " at " << old_node.peer.ip << ":" << old_node.peer.port 
+                  << " to check if alive. Candidate " << node_id_to_hex(candidate_node.id) 
+                  << " waiting to replace if old node fails. (transaction: " << ping_transaction_id << ")");
+    
+    // Store ping verification state and mark old node as being pinged
+    pending_pings_.emplace(ping_transaction_id, PingVerification(candidate_node, old_node, bucket_index));
+    nodes_being_replaced_.insert(old_node.id);
+    
+    // BEP 5: Send ping to the OLD node to verify it's still alive
+    // If old node responds -> keep it, discard candidate
+    // If old node times out -> replace with candidate
+    auto message = KrpcProtocol::create_ping_query(ping_transaction_id, node_id_);
+    send_krpc_message(message, old_node.peer);
+}
+
+void DhtClient::handle_ping_verification_response(const std::string& transaction_id, const NodeId& responder_id, const Peer& responder) {
+    std::lock_guard<std::mutex> ping_lock(pending_pings_mutex_);
+    
+    auto it = pending_pings_.find(transaction_id);
+    if (it != pending_pings_.end()) {
+        const auto& verification = it->second;
+        
+        // BEP 5: We pinged the OLD node to check if it's still alive
+        if (responder_id == verification.old_node.id) {
+            // Calculate RTT
+            auto rtt_duration = std::chrono::steady_clock::now() - verification.ping_sent_at;
+            uint16_t rtt_ms = static_cast<uint16_t>(
+                (std::min)(static_cast<int64_t>(0xfffe),
+                        static_cast<int64_t>(std::chrono::duration_cast<std::chrono::milliseconds>(rtt_duration).count())));
+            
+            // Old node responded - it's still alive! Keep it, discard the candidate.
+            LOG_DHT_DEBUG("Old node " << node_id_to_hex(verification.old_node.id) 
+                          << " responded to ping (rtt=" << rtt_ms << "ms) - keeping it, discarding candidate " 
+                          << node_id_to_hex(verification.candidate_node.id));
+            
+            // Update old node in routing table
+            {
+                std::lock_guard<std::mutex> rt_lock(routing_table_mutex_);
+                auto& bucket = routing_table_[verification.bucket_index];
+                auto node_it = std::find_if(bucket.begin(), bucket.end(),
+                    [&verification](const DhtNode& n) { return n.id == verification.old_node.id; });
+                if (node_it != bucket.end()) {
+                    node_it->mark_success();
+                    node_it->update_rtt(rtt_ms);
+                }
+            }
+            // Candidate is discarded (not added to routing table)
+        } else {
+            LOG_DHT_WARN("Ping verification response from unexpected node " << node_id_to_hex(responder_id) 
+                         << " at " << responder.ip << ":" << responder.port 
+                         << " (expected old node " << node_id_to_hex(verification.old_node.id) << ")");
+        }
+        
+        // Remove tracking entries
+        nodes_being_replaced_.erase(verification.old_node.id);
+        pending_pings_.erase(it);
+    }
+}
+
+void DhtClient::cleanup_stale_ping_verifications() {
+    std::lock_guard<std::mutex> ping_lock(pending_pings_mutex_);
+    
+    auto now = std::chrono::steady_clock::now();
+    auto timeout_threshold = std::chrono::seconds(30);  // 30 second timeout for ping responses
+    
+    auto it = pending_pings_.begin();
+    while (it != pending_pings_.end()) {
+        if (now - it->second.ping_sent_at > timeout_threshold) {
+            const auto& verification = it->second;
+            
+            // BEP 5: Old node didn't respond (timeout) - it's dead, replace with candidate!
+            LOG_DHT_DEBUG("Old node " << node_id_to_hex(verification.old_node.id) 
+                          << " timed out after 30s - replacing with candidate " << node_id_to_hex(verification.candidate_node.id));
+            
+            // Perform the replacement with a fresh candidate
+            DhtNode fresh_candidate = verification.candidate_node;
+            perform_replacement(fresh_candidate, verification.old_node, verification.bucket_index);
+            
+            // Remove tracking entries
+            nodes_being_replaced_.erase(verification.old_node.id);
+            
+            it = pending_pings_.erase(it);
+        } else {
+            ++it;
+        }
+    }
+}
+
+bool DhtClient::perform_replacement(const DhtNode& candidate_node, const DhtNode& node_to_replace, int bucket_index) {
+    std::lock_guard<std::mutex> lock(routing_table_mutex_);
+    
+    auto& bucket = routing_table_[bucket_index];
+    auto it = std::find_if(bucket.begin(), bucket.end(),
+                          [&node_to_replace](const DhtNode& node) {
+                              return node.id == node_to_replace.id;
+                          });
+    
+    if (it != bucket.end()) {
+        LOG_DHT_DEBUG("Replacing old node " << node_id_to_hex(node_to_replace.id) 
+                      << " with " << node_id_to_hex(candidate_node.id) << " in bucket " << bucket_index);
+        *it = candidate_node;
+        return true;
+    } else {
+        LOG_DHT_WARN("Could not find node " << node_id_to_hex(node_to_replace.id) 
+                     << " to replace in bucket " << bucket_index);
+    }
+
+    return false;
+}
+
+// Utility functions implementation
+NodeId string_to_node_id(const std::string& str) {
+    NodeId id;
+    size_t copy_size = (std::min)(str.size(), NODE_ID_SIZE);
+    std::copy(str.begin(), str.begin() + copy_size, id.begin());
+    return id;
+}
+
+std::string node_id_to_string(const NodeId& id) {
+    return std::string(id.begin(), id.end());
+}
+
+NodeId hex_to_node_id(const std::string& hex) {
+    NodeId id;
+    if (hex.size() != NODE_ID_SIZE * 2) {
+        return id;  // Return zero-filled ID on error
+    }
+    
+    for (size_t i = 0; i < NODE_ID_SIZE; ++i) {
+        std::string byte_str = hex.substr(i * 2, 2);
+        id[i] = static_cast<uint8_t>(std::stoul(byte_str, nullptr, 16));
+    }
+    
+    return id;
+}
+
+std::string node_id_to_hex(const NodeId& id) {
+    std::ostringstream oss;
+    oss << std::hex << std::setfill('0');
+    for (uint8_t byte : id) {
+        oss << std::setw(2) << static_cast<int>(byte);
+    }
+    return oss.str();
+}
+
+// Routing table persistence implementation
+bool DhtClient::save_routing_table() {
+    std::lock_guard<std::mutex> lock(routing_table_mutex_);
+    
+    try {
+        nlohmann::json routing_data;
+        routing_data["version"] = 1;
+        routing_data["node_id"] = node_id_to_hex(node_id_);
+        routing_data["saved_at"] = std::chrono::system_clock::now().time_since_epoch().count();
+        
+        nlohmann::json nodes_array = nlohmann::json::array();
+        
+        // Save only good nodes (confirmed with fail_count == 0)
+        size_t saved_count = 0;
+        for (const auto& bucket : routing_table_) {
+            for (const auto& node : bucket) {
+                // Only save confirmed good nodes
+                if (node.confirmed()) {
+                    nlohmann::json node_data;
+                    node_data["id"] = node_id_to_hex(node.id);
+                    node_data["ip"] = node.peer.ip;
+                    node_data["port"] = node.peer.port;
+                    
+                    // Save RTT if known
+                    if (node.rtt != 0xffff) {
+                        node_data["rtt"] = node.rtt;
+                    }
+                    
+                    nodes_array.push_back(node_data);
+                    saved_count++;
+                }
+            }
+        }
+        
+        routing_data["nodes"] = nodes_array;
+        routing_data["count"] = saved_count;
+        
+        // Determine file path
+        std::string file_path;
+        #ifdef TESTING
+            if (port_ == 0) {
+                std::ostringstream oss;
+                oss << "dht_routing_" << this << ".json";
+                file_path = oss.str();
+            } else {
+                file_path = "dht_routing_" + std::to_string(port_) + ".json";
+            }
+        #else
+            file_path = data_directory_ + "/dht_routing_" + std::to_string(port_) + ".json";
+        #endif
+        
+        // Write to file
+        std::ofstream file(file_path);
+        if (!file.is_open()) {
+            LOG_DHT_ERROR("Failed to open routing table file for writing: " << file_path);
+            return false;
+        }
+        
+        file << routing_data.dump(2);
+        file.close();
+        
+        LOG_DHT_DEBUG("Saved " << saved_count << " confirmed nodes to " << file_path);
+        return true;
+        
+    } catch (const std::exception& e) {
+        LOG_DHT_ERROR("Exception while saving routing table: " << e.what());
+        return false;
+    }
+}
+
+bool DhtClient::load_routing_table() {
+    std::lock_guard<std::mutex> lock(routing_table_mutex_);
+    
+    try {
+        // Determine file path
+        std::string file_path;
+        #ifdef TESTING
+            if (port_ == 0) {
+                std::ostringstream oss;
+                oss << "dht_routing_" << this << ".json";
+                file_path = oss.str();
+            } else {
+                file_path = "dht_routing_" + std::to_string(port_) + ".json";
+            }
+        #else
+            file_path = data_directory_ + "/dht_routing_" + std::to_string(port_) + ".json";
+        #endif
+        
+        // Check if file exists
+        std::ifstream file(file_path);
+        if (!file.is_open()) {
+            LOG_DHT_DEBUG("No saved routing table found at " << file_path);
+            return false;
+        }
+        
+        // Parse JSON
+        nlohmann::json routing_data;
+        file >> routing_data;
+        file.close();
+        
+        // Validate format
+        if (!routing_data.contains("version") || !routing_data.contains("nodes")) {
+            LOG_DHT_WARN("Invalid routing table file format");
+            return false;
+        }
+        
+        int version = routing_data["version"];
+        if (version != 1) {
+            LOG_DHT_WARN("Unsupported routing table version: " << version);
+            return false;
+        }
+        
+        // Load nodes
+        const auto& nodes_array = routing_data["nodes"];
+        size_t loaded_count = 0;
+        
+        for (const auto& node_data : nodes_array) {
+            try {
+                std::string node_id_hex = node_data["id"];
+                std::string ip = node_data["ip"];
+                int port = node_data["port"];
+                
+                NodeId node_id = hex_to_node_id(node_id_hex);
+                Peer peer(ip, port);
+                DhtNode node(node_id, peer);
+                
+                // Restore RTT if available
+                if (node_data.contains("rtt")) {
+                    node.rtt = node_data["rtt"];
+                }
+                
+                // Mark as confirmed (fail_count = 0)
+                node.fail_count = 0;
+                
+                // Add to appropriate bucket
+                int bucket_index = get_bucket_index(node.id);
+                auto& bucket = routing_table_[bucket_index];
+                
+                // Check if bucket has space
+                if (bucket.size() < K_BUCKET_SIZE) {
+                    bucket.push_back(node);
+                    loaded_count++;
+                } else {
+                    // Bucket full - try to replace a worse node
+                    auto worst_it = std::max_element(bucket.begin(), bucket.end(),
+                        [](const DhtNode& a, const DhtNode& b) {
+                            return a.is_worse_than(b);
+                        });
+                    
+                    if (worst_it != bucket.end() && worst_it->is_worse_than(node)) {
+                        *worst_it = node;
+                        loaded_count++;
+                    }
+                }
+                
+            } catch (const std::exception& e) {
+                LOG_DHT_WARN("Failed to load node from routing table: " << e.what());
+                continue;
+            }
+        }
+        
+        LOG_DHT_INFO("Loaded " << loaded_count << " nodes from routing table file");
+        return loaded_count > 0;
+        
+    } catch (const std::exception& e) {
+        LOG_DHT_ERROR("Exception while loading routing table: " << e.what());
+        return false;
+    }
+}
+
+void DhtClient::set_data_directory(const std::string& directory) {
+    data_directory_ = directory;
+    if (data_directory_.empty()) {
+        data_directory_ = ".";
+    }
+    LOG_DHT_DEBUG("Data directory set to: " << data_directory_);
+}
+
+} // namespace librats