sf-veritas 0.10.3__cp314-cp314-manylinux_2_28_x86_64.whl

sf_veritas/_sfnetworkhop.c

@@ -0,0 +1,1451 @@
+// sf_veritas/_sfnetworkhop.c
+// Ultra-fast network hop capture with non-blocking producers, no drops,
+// and an HTTP/2 multiplexed libcurl-multi sender.
+
+#define PY_SSIZE_T_CLEAN
+#include <Python.h>
+#include <pthread.h>
+#include <curl/curl.h>
+#include <stdatomic.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
+#include <sys/time.h>
+
+#ifndef SFN_RING_CAP
+#define SFN_RING_CAP 65536  // power-of-two recommended
+#endif
+
+#ifndef SFN_ENDPOINT_CAP
+#define SFN_ENDPOINT_CAP 2048
+#endif
+
+#ifndef SFN_MAX_INFLIGHT
+#define SFN_MAX_INFLIGHT 128   // number of concurrent requests over H2
+#endif
+
+// ---------- Message (three types: fast work, work with bodies, ready body) ----------
+typedef enum {
+    SFN_MSG_WORK,         // raw work item (needs JSON building)
+    SFN_MSG_WORK_BODIES,  // work item with request/response bodies
+    SFN_MSG_BODY          // pre-built body (legacy path)
+} sfn_msg_type;
+
+typedef struct {
+    sfn_msg_type type;
+    union {
+        struct {  // type == SFN_MSG_WORK
+            char *session_id;    // malloced copy
+            size_t session_len;
+            int endpoint_id;
+        } work;
+        struct {  // type == SFN_MSG_WORK_BODIES
+            char *session_id;    // malloced copy
+            size_t session_len;
+            int endpoint_id;
+            char *route;         // malloced JSON-escaped string or NULL (overrides registered route)
+            size_t route_len;
+            char *query_params;  // malloced JSON-escaped string or NULL
+            size_t query_params_len;
+            char *req_headers;   // malloced JSON-escaped string or NULL
+            size_t req_headers_len;
+            char *req_body;      // malloced JSON-escaped string or NULL
+            size_t req_body_len;
+            char *resp_headers;  // malloced JSON-escaped string or NULL
+            size_t resp_headers_len;
+            char *resp_body;     // malloced JSON-escaped string or NULL
+            size_t resp_body_len;
+        } work_bodies;
+        struct {  // type == SFN_MSG_BODY
+            char *body;
+            size_t len;
+        } body;
+    } data;
+} sfn_msg_t;
+
+// ---------- Ring buffer (bounded, cache-friendly) ----------
+static sfn_msg_t *g_ring = NULL;
+static size_t     g_cap  = 0;
+static _Atomic size_t g_head = 0; // consumer index
+static _Atomic size_t g_tail = 0; // producer index
+
+// REMOVED: spinlock for MPMC push (replaced with lock-free CAS below)
+
+// ---------- Overflow (unbounded, no-drop) ----------
+typedef struct sfn_node_t {
+    sfn_msg_t msg;
+    struct sfn_node_t *next;
+} sfn_node_t;
+
+static _Atomic(sfn_node_t*) g_overflow_head = NULL;
+
+// ---------- wake/sleep ----------
+static pthread_mutex_t g_cv_mtx = PTHREAD_MUTEX_INITIALIZER;
+static pthread_cond_t  g_cv     = PTHREAD_COND_INITIALIZER;
+static _Atomic int     g_running = 0;
+
+// Thread pool for concurrent senders (configurable via SF_NETWORKHOP_SENDER_THREADS)
+#define MAX_SENDER_THREADS 16
+static pthread_t g_sender_threads[MAX_SENDER_THREADS];
+static int g_num_sender_threads = 0;
+
+// ---------- libcurl state ----------
+__thread CURLM *g_multi = NULL;     // per-thread multi interface (HTTP/2 multiplexing)
+static CURL  *g_share_template = NULL; // template to clone options from
+static struct curl_slist *g_hdrs = NULL;
+static _Atomic int g_inflight = 0;
+
+// simple easy-handle pool (LIFO)
+typedef struct pool_node_t { CURL *easy; struct pool_node_t *next; } pool_node_t;
+static pool_node_t *g_easy_pool = NULL;
+static pthread_mutex_t g_pool_mtx = PTHREAD_MUTEX_INITIALIZER;
+
+// ---------- config ----------
+static char *g_url = NULL;
+static char *g_query_escaped = NULL;
+static char *g_api_key = NULL;
+static char *g_service_uuid = NULL;
+static int   g_http2 = 0;
+
+// JSON prefix/suffix
+static char *g_json_prefix = NULL;
+static size_t g_json_prefix_len = 0;
+static const char *JSON_SUFFIX = "}}";
+
+// ---------- Endpoint registry ----------
+typedef struct {
+    char  *suffix;      // pre-escaped invariant suffix (ends with ,\"timestampMs\":\")
+    size_t suffix_len;
+    int    in_use;
+} endpoint_entry;
+
+static endpoint_entry g_endpoints[SFN_ENDPOINT_CAP];
+static _Atomic int g_endpoint_count = 0;
+
+// ---------- helpers ----------
+static inline uint64_t now_ms(void) {
+#if defined(CLOCK_REALTIME_COARSE)
+    struct timespec ts;
+    clock_gettime(CLOCK_REALTIME_COARSE, &ts);
+    return ((uint64_t)ts.tv_sec) * 1000ULL + (uint64_t)(ts.tv_nsec / 1000000ULL);
+#else
+    struct timeval tv;
+    gettimeofday(&tv, NULL);
+    return ((uint64_t)tv.tv_sec) * 1000ULL + (uint64_t)(tv.tv_usec / 1000ULL);
+#endif
+}
+
+static char *str_dup(const char *s) {
+    size_t n = strlen(s);
+    char *p = (char*)malloc(n + 1);
+    if (!p) return NULL;
+    memcpy(p, s, n);
+    p[n] = 0;
+    return p;
+}
+
+static char *json_escape(const char *s) {
+    const unsigned char *in = (const unsigned char*)s;
+    size_t extra = 0;
+    for (const unsigned char *p = in; *p; ++p) {
+        switch (*p) {
+            case '\\': case '"': case '\n': case '\r': case '\t': case '\b': case '\f':
+                extra++; break;
+            default:
+                if (*p < 0x20) extra += 5; // \u00XX for other control chars
+        }
+    }
+    size_t inlen = strlen(s);
+    char *out = (char*)malloc(inlen + extra + 1);
+    if (!out) return NULL;
+
+    char *o = out;
+    for (const unsigned char *p = in; *p; ++p) {
+        switch (*p) {
+            case '\\': *o++='\\'; *o++='\\'; break;
+            case '"':  *o++='\\'; *o++='"';  break;
+            case '\n': *o++='\\'; *o++='n';  break;
+            case '\r': *o++='\\'; *o++='r';  break;
+            case '\t': *o++='\\'; *o++='t';  break;
+            case '\b': *o++='\\'; *o++='b';  break;
+            case '\f': *o++='\\'; *o++='f';  break;
+            default:
+                if (*p < 0x20) {
+                    // Unicode escape for rare control chars
+                    static const char hex[] = "0123456789abcdef";
+                    *o++='\\'; *o++='u'; *o++='0'; *o++='0';
+                    *o++=hex[(*p)>>4]; *o++=hex[(*p)&0xF];
+                } else {
+                    *o++ = (char)*p;
+                }
+        }
+    }
+    *o = 0;
+    return out;
+}
+
+static char *json_escape_query(const char *s) {
+    const unsigned char *in = (const unsigned char*)s;
+    size_t extra = 0;
+    for (const unsigned char *p = in; *p; ++p) {
+        switch (*p) {
+            case '\\': case '"': case '\n': case '\r': case '\t': extra++; break;
+            default: break;
+        }
+    }
+    size_t inlen = strlen(s);
+    char *out = (char*)malloc(inlen + extra + 1);
+    if (!out) return NULL;
+    char *o = out;
+    for (const unsigned char *p = in; *p; ++p) {
+        switch (*p) {
+            case '\\': *o++='\\'; *o++='\\'; break;
+            case '"':  *o++='\\'; *o++='"';  break;
+            case '\n': *o++='\\'; *o++='n';  break;
+            case '\r': *o++='\\'; *o++='r';  break;
+            case '\t': *o++='\\'; *o++='t';  break;
+            default: *o++=(char)*p;
+        }
+    }
+    *o=0;
+    return out;
+}
+
+static int build_prefix(void) {
+    const char *p1 = "{\"query\":\"";
+    const char *p2 = "\",\"variables\":{";
+    const char *k1 = "\"apiKey\":\"";
+    const char *k2 = "\",\"serviceUuid\":\"";
+
+    size_t n = strlen(p1) + strlen(g_query_escaped) + strlen(p2)
+             + strlen(k1) + strlen(g_api_key)
+             + strlen(k2) + strlen(g_service_uuid) + 5;
+
+    char *prefix = (char*)malloc(n);
+    if (!prefix) return 0;
+
+    char *o = prefix;
+    o += sprintf(o, "%s%s%s", p1, g_query_escaped, p2);
+    o += sprintf(o, "%s%s", k1, g_api_key);
+    o += sprintf(o, "%s%s\"", k2, g_service_uuid);
+    *o = '\0';
+
+    g_json_prefix = prefix;
+    g_json_prefix_len = (size_t)(o - prefix);
+    return 1;
+}
+
+static size_t json_escape_inline(char *out, const char *in, size_t in_len) {
+    char *o = out;
+    for (size_t i = 0; i < in_len; ++i) {
+        unsigned char c = (unsigned char)in[i];
+        switch (c) {
+            case '\\': *o++='\\'; *o++='\\'; break;
+            case '"':  *o++='\\'; *o++='"';  break;
+            default:
+                if (c < 0x20) {
+                    static const char hex[] = "0123456789abcdef";
+                    *o++='\\'; *o++='u'; *o++='0'; *o++='0';
+                    *o++=hex[c>>4]; *o++=hex[c&0xF];
+                } else {
+                    *o++ = (char)c;
+                }
+        }
+    }
+    return (size_t)(o - out);
+}
+
+static inline size_t max_escaped_size(size_t len) { return len * 6; }
+
+// ---------- Endpoint registry ----------
+static int make_endpoint_suffix(
+    const char *line_e, const char *column_e,
+    const char *name_e, const char *entrypoint_e, const char *route_e,
+    char **out, size_t *out_len
+) {
+    const char *k1 = ",\"line\":\"";
+    const char *k2 = "\",\"column\":\"";
+    const char *k3 = "\",\"name\":\"";
+    const char *k4 = "\",\"entrypoint\":\"";
+    const char *k5 = "\",\"route\":\"";
+    const char *k6 = "\",\"timestampMs\":\"";
+
+    size_t n = strlen(k1)+strlen(line_e)
+             + strlen(k2)+strlen(column_e)
+             + strlen(k3)+strlen(name_e)
+             + strlen(k4)+strlen(entrypoint_e)
+             + strlen(k5)+strlen(route_e)
+             + strlen(k6);
+
+    char *buf = (char*)malloc(n + 1);
+    if (!buf) return 0;
+    char *o = buf;
+    o += sprintf(o, "%s%s%s%s%s%s%s%s%s%s%s",
+                 k1, line_e, k2, column_e, k3, name_e, k4, entrypoint_e, k5, route_e, k6);
+    *o = 0;
+    *out = buf;
+    *out_len = (size_t)(o - buf);
+    return 1;
+}
+
+static int register_endpoint_internal(
+    const char *line, const char *column, const char *name, const char *entrypoint, const char *route
+) {
+    int idx = atomic_fetch_add_explicit(&g_endpoint_count, 1, memory_order_acq_rel);
+    if (idx < 0 || idx >= SFN_ENDPOINT_CAP) return -1;
+
+    char *line_e = json_escape(line);
+    char *col_e  = json_escape(column);
+    char *name_e = json_escape(name);
+    char *ep_e   = json_escape(entrypoint);
+    char *route_e = json_escape(route ? route : "");
+    if (!line_e || !col_e || !name_e || !ep_e || !route_e) {
+        free(line_e); free(col_e); free(name_e); free(ep_e); free(route_e);
+        return -1;
+    }
+    char *suffix = NULL; size_t suffix_len = 0;
+    if (!make_endpoint_suffix(line_e, col_e, name_e, ep_e, route_e, &suffix, &suffix_len)) {
+        free(line_e); free(col_e); free(name_e); free(ep_e); free(route_e);
+        return -1;
+    }
+    free(line_e); free(col_e); free(name_e); free(ep_e); free(route_e);
+
+    g_endpoints[idx].suffix = suffix;
+    g_endpoints[idx].suffix_len = suffix_len;
+    g_endpoints[idx].in_use = 1;
+    return idx;
+}
+
+// ---------- message helpers ----------
+static inline void msg_free(sfn_msg_t *msg) {
+    if (!msg) return;
+    if (msg->type == SFN_MSG_WORK) {
+        free(msg->data.work.session_id);
+    } else if (msg->type == SFN_MSG_WORK_BODIES) {
+        free(msg->data.work_bodies.session_id);
+        free(msg->data.work_bodies.route);
+        free(msg->data.work_bodies.query_params);
+        free(msg->data.work_bodies.req_headers);
+        free(msg->data.work_bodies.req_body);
+        free(msg->data.work_bodies.resp_headers);
+        free(msg->data.work_bodies.resp_body);
+    } else if (msg->type == SFN_MSG_BODY) {
+        free(msg->data.body.body);
+    }
+}
+
+// ---------- overflow ops ----------
+static inline void overflow_push(sfn_msg_t msg) {
+    sfn_node_t *n = (sfn_node_t*)malloc(sizeof(sfn_node_t));
+    if (!n) { msg_free(&msg); return; }
+    n->msg = msg;
+    sfn_node_t *old = atomic_load_explicit(&g_overflow_head, memory_order_relaxed);
+    do { n->next = old; }
+    while (!atomic_compare_exchange_weak_explicit(
+        &g_overflow_head, &old, n, memory_order_release, memory_order_relaxed));
+
+    pthread_mutex_lock(&g_cv_mtx);
+    pthread_cond_signal(&g_cv);
+    pthread_mutex_unlock(&g_cv_mtx);
+}
+
+static inline sfn_node_t* overflow_pop_all(void) {
+    return atomic_exchange_explicit(&g_overflow_head, NULL, memory_order_acq_rel);
+}
+
+static inline void overflow_free_list(sfn_node_t* list) {
+    while (list) {
+        sfn_node_t* next = list->next;
+        msg_free(&list->msg);
+        free(list);
+        list = next;
+    }
+}
+
+// ---------- ring ops ----------
+static inline size_t ring_count(void) {
+    size_t h = atomic_load_explicit(&g_head, memory_order_acquire);
+    size_t t = atomic_load_explicit(&g_tail, memory_order_acquire);
+    return t - h;
+}
+static inline int ring_empty(void) { return ring_count() == 0; }
+
+// PERFORMANCE: Lock-free CAS-based push (removes global contention point)
+// Many producers can make forward progress concurrently without spinning
+static int ring_try_push(sfn_msg_t msg) {
+    for (;;) {
+        size_t t = atomic_load_explicit(&g_tail, memory_order_relaxed);
+        size_t h = atomic_load_explicit(&g_head, memory_order_acquire);
+
+        if ((t - h) >= g_cap) {
+            return 0; // full
+        }
+
+        // Try to claim slot t via CAS
+        if (atomic_compare_exchange_weak_explicit(
+                &g_tail, &t, t + 1, memory_order_acq_rel, memory_order_relaxed)) {
+
+            int was_empty = (h == t);
+            size_t idx = t % g_cap;
+            g_ring[idx] = msg; // single writer for this idx (we own it after CAS succeeds)
+
+            if (was_empty) {
+                pthread_mutex_lock(&g_cv_mtx);
+                pthread_cond_signal(&g_cv);
+                pthread_mutex_unlock(&g_cv_mtx);
+#if LIBCURL_VERSION_NUM >= 0x073E00  // 7.62.0+
+                // Also wake the multi loop if it's sleeping (immediate response)
+                if (g_multi) curl_multi_wakeup(g_multi);
+#endif
+            }
+            return 1;
+        }
+        // else: lost race, retry
+    }
+}
+
+static int ring_pop(sfn_msg_t *out_msg) {
+    size_t h = atomic_load_explicit(&g_head, memory_order_relaxed);
+    size_t t = atomic_load_explicit(&g_tail, memory_order_acquire);
+    if (h == t) return 0;
+    size_t idx = h % g_cap;
+    *out_msg = g_ring[idx];
+    // Clear for safety
+    g_ring[idx].type = SFN_MSG_BODY;
+    g_ring[idx].data.body.body = NULL;
+    g_ring[idx].data.body.len = 0;
+    atomic_store_explicit(&g_head, h + 1, memory_order_release);
+    return 1;
+}
+
+// ---------- curl callbacks ----------
+static size_t _sink_write(char *ptr, size_t size, size_t nmemb, void *userdata) {
+    (void)ptr; (void)userdata; return size * nmemb;
+}
+static size_t _sink_header(char *ptr, size_t size, size_t nmemb, void *userdata) {
+    (void)ptr; (void)userdata; return size * nmemb;
+}
+
+// ---------- easy-handle pool ----------
+static CURL* pool_acquire_easy(void) {
+    pthread_mutex_lock(&g_pool_mtx);
+    pool_node_t *node = g_easy_pool;
+    if (node) g_easy_pool = node->next;
+    pthread_mutex_unlock(&g_pool_mtx);
+    if (node) {
+        CURL *e = node->easy;
+        free(node);
+        // PERFORMANCE: NO curl_easy_reset() - keep invariants intact!
+        // This eliminates option reconfiguration churn (libcurl connection reuse becomes cheap)
+        // Only per-message options (POSTFIELDS/SIZE/PRIVATE) are set in multi_add_message()
+        return e;
+    }
+    // new handle
+    CURL *e = curl_easy_init();
+    if (!e) return NULL;
+    curl_easy_setopt(e, CURLOPT_URL, g_url);
+    curl_easy_setopt(e, CURLOPT_HTTPHEADER, g_hdrs);
+    curl_easy_setopt(e, CURLOPT_WRITEFUNCTION, _sink_write);
+    curl_easy_setopt(e, CURLOPT_HEADERFUNCTION, _sink_header);
+#ifdef CURL_HTTP_VERSION_2TLS
+    if (g_http2) curl_easy_setopt(e, CURLOPT_HTTP_VERSION, CURL_HTTP_VERSION_2TLS);
+#endif
+    curl_easy_setopt(e, CURLOPT_TCP_KEEPALIVE, 1L);
+    curl_easy_setopt(e, CURLOPT_NOSIGNAL, 1L);
+#ifdef TCP_NODELAY
+    curl_easy_setopt(e, CURLOPT_TCP_NODELAY, 1L);
+#endif
+    return e;
+}
+
+static void pool_release_easy(CURL *e) {
+    if (!e) return;
+    pool_node_t *node = (pool_node_t*)malloc(sizeof(pool_node_t));
+    if (!node) { curl_easy_cleanup(e); return; }
+    node->easy = e;
+    pthread_mutex_lock(&g_pool_mtx);
+    node->next = g_easy_pool;
+    g_easy_pool = node;
+    pthread_mutex_unlock(&g_pool_mtx);
+}
+
+// ---------- JSON building (on background thread) ----------
+static char* build_body_from_work(const char *session_id, size_t session_len, int endpoint_id, size_t *out_len) {
+    if (endpoint_id < 0 || endpoint_id >= SFN_ENDPOINT_CAP) return NULL;
+    if (!g_endpoints[endpoint_id].in_use) return NULL;
+
+    // NOW get the timestamp (on background thread, not request path!)
+    uint64_t tms = now_ms();
+    char ts_buf[32];
+    int ts_len = snprintf(ts_buf, sizeof(ts_buf), "%llu", (unsigned long long)tms);
+
+    size_t sess_frag_max = 14 + max_escaped_size(session_len) + 1;
+    size_t total_max = g_json_prefix_len + sess_frag_max
+                     + g_endpoints[endpoint_id].suffix_len
+                     + (size_t)ts_len + 2;
+
+    char *body = (char*)malloc(total_max + 1);
+    if (!body) return NULL;
+
+    char *o = body;
+    memcpy(o, g_json_prefix, g_json_prefix_len); o += g_json_prefix_len;
+    memcpy(o, ",\"sessionId\":\"", 14); o += 14;
+    o += json_escape_inline(o, session_id, session_len); *o++='"';
+    memcpy(o, g_endpoints[endpoint_id].suffix, g_endpoints[endpoint_id].suffix_len);
+    o += g_endpoints[endpoint_id].suffix_len;
+    memcpy(o, ts_buf, (size_t)ts_len); o += ts_len;
+    *o++ = '"'; memcpy(o, JSON_SUFFIX, 2); o += 2; *o = 0;
+
+    *out_len = (size_t)(o - body);
+    return body;
+}
+
+// Build JSON body with request/response headers and bodies (on background thread)
+static char* build_body_from_work_with_bodies(
+    const char *session_id, size_t session_len, int endpoint_id,
+    const char *route, size_t route_len,
+    const char *query_params, size_t query_params_len,
+    const char *req_headers, size_t req_headers_len,
+    const char *req_body, size_t req_body_len,
+    const char *resp_headers, size_t resp_headers_len,
+    const char *resp_body, size_t resp_body_len,
+    size_t *out_len
+) {
+    if (endpoint_id < 0 || endpoint_id >= SFN_ENDPOINT_CAP) return NULL;
+    if (!g_endpoints[endpoint_id].in_use) return NULL;
+
+    // Get timestamp (on background thread)
+    uint64_t tms = now_ms();
+    char ts_buf[32];
+    int ts_len = snprintf(ts_buf, sizeof(ts_buf), "%llu", (unsigned long long)tms);
+
+    // Calculate max size (headers/bodies are already JSON-escaped)
+    size_t sess_frag_max = 14 + max_escaped_size(session_len) + 1;
+    size_t total_max = g_json_prefix_len + sess_frag_max
+                     + g_endpoints[endpoint_id].suffix_len
+                     + (size_t)ts_len;
+
+    // Add space for route/query override fields (already escaped)
+    if (route) total_max += 12 + route_len;  // ",\"route\":\""
+    if (query_params) total_max += 18 + query_params_len;  // ",\"queryParams\":\""
+
+    // Add space for request/response fields (already escaped, just need quotes and keys)
+    if (req_headers) total_max += 20 + req_headers_len;  // ",\"requestHeaders\":\""
+    if (req_body) total_max += 17 + req_body_len;        // ",\"requestBody\":\""
+    if (resp_headers) total_max += 21 + resp_headers_len; // ",\"responseHeaders\":\""
+    if (resp_body) total_max += 18 + resp_body_len;      // ",\"responseBody\":\""
+    total_max += 10; // closing quotes and braces
+
+    char *body = (char*)malloc(total_max + 1);
+    if (!body) return NULL;
+
+    char *o = body;
+    memcpy(o, g_json_prefix, g_json_prefix_len); o += g_json_prefix_len;
+    memcpy(o, ",\"sessionId\":\"", 14); o += 14;
+    o += json_escape_inline(o, session_id, session_len); *o++='"';
+    memcpy(o, g_endpoints[endpoint_id].suffix, g_endpoints[endpoint_id].suffix_len);
+    o += g_endpoints[endpoint_id].suffix_len;
+    memcpy(o, ts_buf, (size_t)ts_len); o += ts_len;
+    *o++ = '"';
+
+    // Add route override if present (already JSON-escaped) - this overrides the registered route in suffix
+    if (route && route_len > 0) {
+        memcpy(o, ",\"route\":\"", 10); o += 10;
+        memcpy(o, route, route_len); o += route_len;
+        *o++ = '"';
+    }
+
+    // Add query params if present (already JSON-escaped)
+    if (query_params && query_params_len > 0) {
+        memcpy(o, ",\"queryParams\":\"", 16); o += 16;
+        memcpy(o, query_params, query_params_len); o += query_params_len;
+        *o++ = '"';
+    }
+
+    // Add request/response fields if present (already JSON-escaped)
+    if (req_headers && req_headers_len > 0) {
+        memcpy(o, ",\"requestHeaders\":\"", 19); o += 19;
+        memcpy(o, req_headers, req_headers_len); o += req_headers_len;
+        *o++ = '"';
+    }
+    if (req_body && req_body_len > 0) {
+        memcpy(o, ",\"requestBody\":\"", 16); o += 16;
+        memcpy(o, req_body, req_body_len); o += req_body_len;
+        *o++ = '"';
+    }
+    if (resp_headers && resp_headers_len > 0) {
+        memcpy(o, ",\"responseHeaders\":\"", 20); o += 20;
+        memcpy(o, resp_headers, resp_headers_len); o += resp_headers_len;
+        *o++ = '"';
+    }
+    if (resp_body && resp_body_len > 0) {
+        memcpy(o, ",\"responseBody\":\"", 17); o += 17;
+        memcpy(o, resp_body, resp_body_len); o += resp_body_len;
+        *o++ = '"';
+    }
+
+    memcpy(o, JSON_SUFFIX, 2); o += 2; *o = 0;
+
+    *out_len = (size_t)(o - body);
+    return body;
+}
+
+// ---------- multi helpers ----------
+static void multi_add_message(char *body, size_t len) {
+    CURL *e = pool_acquire_easy();
+    if (!e) { free(body); return; }
+
+    curl_easy_setopt(e, CURLOPT_POSTFIELDS, body);
+    curl_easy_setopt(e, CURLOPT_POSTFIELDSIZE_LARGE, (curl_off_t)len);
+    // keep pointer to free later
+    curl_easy_setopt(e, CURLOPT_PRIVATE, body);
+
+    CURLMcode mc = curl_multi_add_handle(g_multi, e);
+    if (mc != CURLM_OK) {
+        pool_release_easy(e);
+        free(body);
+        return;
+    }
+    atomic_fetch_add_explicit(&g_inflight, 1, memory_order_acq_rel);
+
+#if LIBCURL_VERSION_NUM >= 0x073E00  // 7.62.0+
+    // PERFORMANCE: Wake multi loop immediately (avoid 50ms sleep tail latency)
+    curl_multi_wakeup(g_multi);
+#endif
+}
+
+static void process_msg(sfn_msg_t *msg) {
+    if (msg->type == SFN_MSG_WORK) {
+        // Build JSON from work item (expensive work happens HERE, not in request path)
+        size_t len = 0;
+        char *body = build_body_from_work(
+            msg->data.work.session_id,
+            msg->data.work.session_len,
+            msg->data.work.endpoint_id,
+            &len
+        );
+        if (body) {
+            multi_add_message(body, len);
+        }
+        free(msg->data.work.session_id); // done with session_id
+    } else if (msg->type == SFN_MSG_WORK_BODIES) {
+        // Build JSON with request/response bodies (expensive work on background thread)
+        size_t len = 0;
+        char *body = build_body_from_work_with_bodies(
+            msg->data.work_bodies.session_id,
+            msg->data.work_bodies.session_len,
+            msg->data.work_bodies.endpoint_id,
+            msg->data.work_bodies.route,
+            msg->data.work_bodies.route_len,
+            msg->data.work_bodies.query_params,
+            msg->data.work_bodies.query_params_len,
+            msg->data.work_bodies.req_headers,
+            msg->data.work_bodies.req_headers_len,
+            msg->data.work_bodies.req_body,
+            msg->data.work_bodies.req_body_len,
+            msg->data.work_bodies.resp_headers,
+            msg->data.work_bodies.resp_headers_len,
+            msg->data.work_bodies.resp_body,
+            msg->data.work_bodies.resp_body_len,
+            &len
+        );
+        if (body) {
+            multi_add_message(body, len);
+        }
+        // Free all work_bodies fields
+        free(msg->data.work_bodies.session_id);
+        free(msg->data.work_bodies.route);
+        free(msg->data.work_bodies.query_params);
+        free(msg->data.work_bodies.req_headers);
+        free(msg->data.work_bodies.req_body);
+        free(msg->data.work_bodies.resp_headers);
+        free(msg->data.work_bodies.resp_body);
+    } else if (msg->type == SFN_MSG_BODY) {
+        // Already built (legacy path)
+        multi_add_message(msg->data.body.body, msg->data.body.len);
+    }
+}
+
+static void multi_pump(void) {
+    int running = 0;
+    // Drive state machine
+    curl_multi_perform(g_multi, &running);
+
+    // Reap completions
+    int msgs = 0;
+    CURLMsg *msg;
+    while ((msg = curl_multi_info_read(g_multi, &msgs))) {
+        if (msg->msg == CURLMSG_DONE) {
+            CURL *e = msg->easy_handle;
+            char *body = NULL;
+            curl_easy_getinfo(e, CURLINFO_PRIVATE, &body);
+            curl_multi_remove_handle(g_multi, e);
+            pool_release_easy(e);
+            free(body);
+            atomic_fetch_sub_explicit(&g_inflight, 1, memory_order_acq_rel);
+        }
+    }
+}
+
+// ---------- Network capture suppression ----------
+// Define the telemetry guard to prevent recursive capture
+// Each C extension is compiled separately, so we need our own thread-local copy
+__thread int g_in_telemetry_send = 0;
+
+// ---------- pthread cleanup handler for sender threads ----------
+static void sender_cleanup(void *arg) {
+    (void)arg;
+    if (g_multi) {
+        curl_multi_cleanup(g_multi);
+        g_multi = NULL;
+    }
+}
+
+// ---------- sender thread (multi + H2 multiplex) ----------
+static void *sender_main(void *arg) {
+    (void)arg;
+
+    // CRITICAL: Set telemetry guard for this thread to prevent _sfteepreload.c from capturing our traffic
+    // This sender thread only sends telemetry, so ALL its network traffic should be suppressed
+    g_in_telemetry_send = 1;
+
+    // Initialize per-thread curl_multi handle
+    g_multi = curl_multi_init();
+    if (!g_multi) return NULL;
+#ifdef CURLPIPE_MULTIPLEX
+    curl_multi_setopt(g_multi, CURLMOPT_PIPELINING, CURLPIPE_MULTIPLEX);
+#endif
+    curl_multi_setopt(g_multi, CURLMOPT_MAX_TOTAL_CONNECTIONS, 1L);
+    curl_multi_setopt(g_multi, CURLMOPT_MAX_HOST_CONNECTIONS, 1L);
+
+    // Register cleanup handler
+    pthread_cleanup_push(sender_cleanup, NULL);
+
+    while (atomic_load(&g_running)) {
+        // Fill up inflight up to SFN_MAX_INFLIGHT
+        while (atomic_load_explicit(&g_inflight, memory_order_acquire) < SFN_MAX_INFLIGHT) {
+            sfn_msg_t msg;
+            if (ring_pop(&msg)) {
+                process_msg(&msg);
+                continue;
+            }
+            // drain overflow into ring-ish order (reverse the LIFO)
+            sfn_node_t *list = overflow_pop_all();
+            if (list) {
+                // reverse to approx FIFO
+                sfn_node_t *rev = NULL;
+                while (list) { sfn_node_t *n = list->next; list->next = rev; rev = list; list = n; }
+                while (rev && atomic_load_explicit(&g_inflight, memory_order_acquire) < SFN_MAX_INFLIGHT) {
+                    sfn_node_t *n = rev->next;
+                    process_msg(&rev->msg);
+                    free(rev);
+                    rev = n;
+                }
+                // any leftover (shouldn't happen often) goes back to overflow
+                while (rev) { sfn_node_t *n = rev->next; overflow_push(rev->msg); free(rev); rev = n; }
+                continue;
+            }
+            break; // nothing to send
+        }
+
+        // Drive transfers & poll
+        multi_pump();
+
+        // Sleep a bit if idle; otherwise use multi_poll for efficient wake-up
+        if (atomic_load_explicit(&g_inflight, memory_order_acquire) == 0 &&
+            ring_empty() &&
+            atomic_load_explicit(&g_overflow_head, memory_order_acquire) == NULL) {
+            pthread_mutex_lock(&g_cv_mtx);
+            if (atomic_load_explicit(&g_inflight, memory_order_acquire) == 0 &&
+                ring_empty() &&
+                atomic_load_explicit(&g_overflow_head, memory_order_acquire) == NULL &&
+                atomic_load(&g_running)) {
+                pthread_cond_wait(&g_cv, &g_cv_mtx);
+            }
+            pthread_mutex_unlock(&g_cv_mtx);
+        } else {
+            // PERFORMANCE: Short poll wait (5ms) - curl_multi_wakeup handles immediate nudges
+            // This removes scheduler-scale latency without busy-spinning
+            int numfds = 0;
+#if LIBCURL_VERSION_NUM >= 0x074200  // curl >= 7.66.0
+            curl_multi_poll(g_multi, NULL, 0, 5, &numfds);  // 5ms guard (was 50ms)
+#else
+            curl_multi_wait(g_multi, NULL, 0, 5, &numfds);
+#endif
+        }
+    }
+    // final drain
+    while (atomic_load_explicit(&g_inflight, memory_order_acquire) > 0) {
+        multi_pump();
+        int nfds = 0;
+#if LIBCURL_VERSION_NUM >= 0x074200  // curl >= 7.66.0
+        curl_multi_poll(g_multi, NULL, 0, 10, &nfds);
+#else
+        curl_multi_wait(g_multi, NULL, 0, 10, &nfds);
+#endif
+    }
+
+    pthread_cleanup_pop(1);
+    return NULL;
+}
+
+// ---------- Python API ----------
+static PyObject *py_init(PyObject *self, PyObject *args, PyObject *kw) {
+    const char *url, *query, *api_key, *service_uuid;
+    int http2 = 0;
+    static char *kwlist[] = {"url","query","api_key","service_uuid","http2", NULL};
+    if (!PyArg_ParseTupleAndKeywords(args, kw, "ssss|i",
+        kwlist, &url, &query, &api_key, &service_uuid, &http2)) {
+        Py_RETURN_FALSE;
+    }
+    if (g_running) Py_RETURN_TRUE;
+
+    g_url = str_dup(url);
+    g_query_escaped = json_escape_query(query);
+    g_api_key = str_dup(api_key);
+    g_service_uuid = str_dup(service_uuid);
+    g_http2 = http2 ? 1 : 0;
+    if (!g_url || !g_query_escaped || !g_api_key || !g_service_uuid) {
+        Py_RETURN_FALSE;
+    }
+    if (!build_prefix()) { Py_RETURN_FALSE; }
+
+    g_cap = SFN_RING_CAP;
+    g_ring = (sfn_msg_t*)calloc(g_cap, sizeof(sfn_msg_t));
+    if (!g_ring) { Py_RETURN_FALSE; }
+
+    curl_global_init(CURL_GLOBAL_DEFAULT);
+
+    // common headers (REMOVED: X-Sf3-IsTelemetryMessage - no longer needed with guard flag)
+    g_hdrs = NULL;
+    g_hdrs = curl_slist_append(g_hdrs, "Content-Type: application/json");
+    g_hdrs = curl_slist_append(g_hdrs, "Expect:");  // PERFORMANCE: Disable 100-continue (avoids 200ms stalls)
+
+    // NOTE: g_multi is now initialized per-thread in sender_main()
+    // This allows multiple sender threads each with their own curl_multi instance
+
+    // a template easy to copy options from (kept as reference; not added to multi)
+    g_share_template = curl_easy_init();
+    if (!g_share_template) { Py_RETURN_FALSE; }
+    curl_easy_setopt(g_share_template, CURLOPT_URL, g_url);
+    curl_easy_setopt(g_share_template, CURLOPT_HTTPHEADER, g_hdrs);
+    curl_easy_setopt(g_share_template, CURLOPT_WRITEFUNCTION, _sink_write);
+    curl_easy_setopt(g_share_template, CURLOPT_HEADERFUNCTION, _sink_header);
+#ifdef CURL_HTTP_VERSION_2TLS
+    if (g_http2) curl_easy_setopt(g_share_template, CURLOPT_HTTP_VERSION, CURL_HTTP_VERSION_2TLS);
+#endif
+    curl_easy_setopt(g_share_template, CURLOPT_TCP_KEEPALIVE, 1L);
+    curl_easy_setopt(g_share_template, CURLOPT_NOSIGNAL, 1L);
+#ifdef TCP_NODELAY
+    curl_easy_setopt(g_share_template, CURLOPT_TCP_NODELAY, 1L);
+#endif
+
+    // Parse SF_NETWORKHOP_SENDER_THREADS environment variable (default: 2, max: 16)
+    // NetworkHop expected to have decent volume, so default to 2 threads
+    const char *num_threads_env = getenv("SF_NETWORKHOP_SENDER_THREADS");
+    g_num_sender_threads = num_threads_env ? atoi(num_threads_env) : 2;
+    if (g_num_sender_threads < 1) g_num_sender_threads = 1;
+    if (g_num_sender_threads > MAX_SENDER_THREADS) g_num_sender_threads = MAX_SENDER_THREADS;
+
+    atomic_store(&g_running, 1);
+
+    // Start thread pool
+    for (int i = 0; i < g_num_sender_threads; i++) {
+        if (pthread_create(&g_sender_threads[i], NULL, sender_main, NULL) != 0) {
+            atomic_store(&g_running, 0);
+            // Join any threads that were already created
+            for (int j = 0; j < i; j++) {
+                pthread_join(g_sender_threads[j], NULL);
+            }
+            Py_RETURN_FALSE;
+        }
+    }
+
+    Py_RETURN_TRUE;
+}
+
+static PyObject *py_register_endpoint(PyObject *self, PyObject *args, PyObject *kw) {
+    const char *line, *column, *name, *entrypoint, *route = NULL;
+    static char *kwlist[] = {"line","column","name","entrypoint","route", NULL};
+    if (!PyArg_ParseTupleAndKeywords(args, kw, "ssss|z", kwlist,
+                                     &line, &column, &name, &entrypoint, &route)) {
+        return PyLong_FromLong(-1);
+    }
+    if (!g_running || g_json_prefix == NULL) {
+        return PyLong_FromLong(-1);
+    }
+    int idx = register_endpoint_internal(line, column, name, entrypoint, route);
+    return PyLong_FromLong(idx);
+}
+
+// generic (kept); producer path is GIL-free
+static PyObject *py_networkhop(PyObject *self, PyObject *args, PyObject *kw) {
+    const char *session_id, *line, *column, *name, *entrypoint;
+    Py_ssize_t session_len = 0, line_len = 0, column_len = 0, name_len = 0, entrypoint_len = 0;
+    static char *kwlist[] = {"session_id","line","column","name","entrypoint", NULL};
+    if (!PyArg_ParseTupleAndKeywords(args, kw, "s#s#s#s#s#", kwlist,
+                                     &session_id, &session_len,
+                                     &line, &line_len,
+                                     &column, &column_len,
+                                     &name, &name_len,
+                                     &entrypoint, &entrypoint_len)) {
+        Py_RETURN_NONE;
+    }
+    if (!g_running || g_json_prefix == NULL) Py_RETURN_NONE;
+
+    char *body = NULL; size_t len = 0;
+
+    Py_BEGIN_ALLOW_THREADS
+    // Build the full body
+    {
+        uint64_t tms = now_ms();
+        char ts_buf[32];
+        int ts_len = snprintf(ts_buf, sizeof(ts_buf), "%llu", (unsigned long long)tms);
+
+        size_t sess_frag_max = 14 + max_escaped_size((size_t)session_len) + 1;
+        size_t max_len = g_json_prefix_len + sess_frag_max + 200
+                       + max_escaped_size((size_t)line_len)
+                       + max_escaped_size((size_t)column_len)
+                       + max_escaped_size((size_t)name_len)
+                       + max_escaped_size((size_t)entrypoint_len)
+                       + (size_t)ts_len + 2;
+
+        body = (char*)malloc(max_len + 1);
+        if (body) {
+            char *o = body;
+            memcpy(o, g_json_prefix, g_json_prefix_len); o += g_json_prefix_len;
+            memcpy(o, ",\"sessionId\":\"", 14); o += 14;
+            o += json_escape_inline(o, session_id, (size_t)session_len); *o++='"';
+
+            memcpy(o, ",\"line\":\"", 10); o += 10; o += json_escape_inline(o, line, (size_t)line_len);
+            memcpy(o, "\",\"column\":\"", 12); o += 12; o += json_escape_inline(o, column, (size_t)column_len);
+            memcpy(o, "\",\"name\":\"", 10); o += 10; o += json_escape_inline(o, name, (size_t)name_len);
+            memcpy(o, "\",\"entrypoint\":\"", 16); o += 16; o += json_escape_inline(o, entrypoint, (size_t)entrypoint_len);
+
+            memcpy(o, "\",\"timestampMs\":\"", 17); o += 17;
+            memcpy(o, ts_buf, (size_t)ts_len); o += ts_len;
+            *o++ = '"'; memcpy(o, JSON_SUFFIX, 2); o += 2; *o = 0;
+
+            len = (size_t)(o - body);
+
+            sfn_msg_t msg = { .type = SFN_MSG_BODY, .data.body = { .body = body, .len = len } };
+            if (!ring_try_push(msg)) overflow_push(msg);
+        }
+    }
+    Py_END_ALLOW_THREADS
+
+    Py_RETURN_NONE;
+}
+
+// ULTRA-FAST PATH: Zero-copy - just queue the pointer, background thread copies
+// This achieves TRUE async: only atomic queue operation on hot path (~1µs)
+static PyObject *py_networkhop_fast(PyObject *self, PyObject *args, PyObject *kw) {
+    const char *session_id; Py_ssize_t session_len = 0; int endpoint_id = -1;
+    static char *kwlist[] = {"session_id","endpoint_id", NULL};
+    if (!PyArg_ParseTupleAndKeywords(args, kw, "s#i", kwlist,
+                                     &session_id, &session_len, &endpoint_id)) {
+        Py_RETURN_NONE;
+    }
+    if (!g_running || g_json_prefix == NULL) Py_RETURN_NONE;
+    if (endpoint_id < 0 || endpoint_id >= SFN_ENDPOINT_CAP) Py_RETURN_NONE;
+    if (!g_endpoints[endpoint_id].in_use) Py_RETURN_NONE;
+
+    // CRITICAL OPTIMIZATION: Do the malloc+memcpy in a SINGLE atomic operation
+    // to minimize time holding the GIL. This is faster than releasing GIL,
+    // doing malloc, then reacquiring.
+    size_t len = (size_t)session_len;
+    char *sid_copy = (char*)malloc(len + 1);
+    if (!sid_copy) Py_RETURN_NONE;
+
+    memcpy(sid_copy, session_id, len);
+    sid_copy[len] = 0;
+
+    sfn_msg_t msg = {
+        .type = SFN_MSG_WORK,
+        .data.work = {
+            .session_id = sid_copy,
+            .session_len = len,
+            .endpoint_id = endpoint_id
+        }
+    };
+
+    // Release GIL for queue operation (this is the only blocking part)
+    int pushed = 0;
+    Py_BEGIN_ALLOW_THREADS
+    pushed = ring_try_push(msg);
+    if (!pushed) overflow_push(msg);
+    Py_END_ALLOW_THREADS
+
+    Py_RETURN_NONE;
+}
+
+// Helper: concatenate dict of header key-value pairs or list of bytes chunks (with GIL held)
+// Returns malloced JSON-escaped string
+static char* extract_and_escape_data(PyObject *obj, size_t *out_len) {
+    if (!obj || obj == Py_None) {
+        *out_len = 0;
+        return NULL;
+    }
+
+    // Handle dict (headers) - format as JSON object string
+    if (PyDict_Check(obj)) {
+        // Build JSON object: {"key":"value","key2":"value2"}
+        PyObject *items = PyDict_Items(obj);
+        if (!items) { *out_len = 0; return NULL; }
+
+        Py_ssize_t nitems = PyList_Size(items);
+        if (nitems == 0) {
+            Py_DECREF(items);
+            *out_len = 0;
+            return NULL;
+        }
+
+        // Build JSON object (will be double-escaped: once here, once when inserted into variables)
+        // Estimate size: {"key":"value","key2":"value2"}
+        size_t total = 2; // {}
+        for (Py_ssize_t i = 0; i < nitems; i++) {
+            PyObject *pair = PyList_GetItem(items, i);
+            PyObject *key = PyTuple_GetItem(pair, 0);
+            PyObject *val = PyTuple_GetItem(pair, 1);
+            const char *key_str = PyUnicode_AsUTF8(key);
+            const char *val_str = PyUnicode_AsUTF8(val);
+            if (key_str && val_str) {
+                // "key":"value", (worst case: double length for escaping)
+                total += strlen(key_str) * 2 + strlen(val_str) * 2 + 6; // quotes, colon, comma
+            }
+        }
+
+        char *result = (char*)malloc(total * 2 + 1); // *2 for extra safety
+        if (!result) {
+            Py_DECREF(items);
+            *out_len = 0;
+            return NULL;
+        }
+
+        char *o = result;
+        *o++ = '{';
+        int first = 1;
+        for (Py_ssize_t i = 0; i < nitems; i++) {
+            PyObject *pair = PyList_GetItem(items, i);
+            PyObject *key = PyTuple_GetItem(pair, 0);
+            PyObject *val = PyTuple_GetItem(pair, 1);
+
+            // Convert key to string (keys are always strings in JSON)
+            const char *key_str = NULL;
+            PyObject *key_str_obj = NULL;
+            if (PyUnicode_Check(key)) {
+                key_str = PyUnicode_AsUTF8(key);
+            } else {
+                key_str_obj = PyObject_Str(key);
+                if (key_str_obj) {
+                    key_str = PyUnicode_AsUTF8(key_str_obj);
+                }
+            }
+
+            if (!key_str) {
+                Py_XDECREF(key_str_obj);
+                if (PyErr_Occurred()) PyErr_Clear();
+                continue;
+            }
+
+            if (!first) *o++ = ',';
+            first = 0;
+
+            // Escape key (always quoted)
+            *o++ = '"';
+            char *key_esc = json_escape(key_str);
+            if (key_esc) {
+                size_t klen = strlen(key_esc);
+                memcpy(o, key_esc, klen);
+                o += klen;
+                free(key_esc);
+            }
+            *o++ = '"';
+            *o++ = ':';
+
+            // Handle value based on type (preserve JSON types)
+            if (val == Py_None) {
+                // null
+                memcpy(o, "null", 4);
+                o += 4;
+            } else if (PyBool_Check(val)) {
+                // true or false
+                if (val == Py_True) {
+                    memcpy(o, "true", 4);
+                    o += 4;
+                } else {
+                    memcpy(o, "false", 5);
+                    o += 5;
+                }
+            } else if (PyLong_Check(val)) {
+                // integer (no quotes)
+                long num = PyLong_AsLong(val);
+                if (num == -1 && PyErr_Occurred()) {
+                    PyErr_Clear();
+                    memcpy(o, "null", 4);
+                    o += 4;
+                } else {
+                    char buf[32];
+                    int len = snprintf(buf, sizeof(buf), "%ld", num);
+                    memcpy(o, buf, len);
+                    o += len;
+                }
+            } else if (PyFloat_Check(val)) {
+                // float (no quotes)
+                double num = PyFloat_AsDouble(val);
+                if (num == -1.0 && PyErr_Occurred()) {
+                    PyErr_Clear();
+                    memcpy(o, "null", 4);
+                    o += 4;
+                } else {
+                    char buf[64];
+                    int len = snprintf(buf, sizeof(buf), "%.17g", num);
+                    memcpy(o, buf, len);
+                    o += len;
+                }
+            } else {
+                // String or other type - convert to string and quote
+                const char *val_str = NULL;
+                PyObject *val_str_obj = NULL;
+                if (PyUnicode_Check(val)) {
+                    val_str = PyUnicode_AsUTF8(val);
+                } else {
+                    val_str_obj = PyObject_Str(val);
+                    if (val_str_obj) {
+                        val_str = PyUnicode_AsUTF8(val_str_obj);
+                    }
+                }
+
+                if (val_str) {
+                    *o++ = '"';
+                    char *val_esc = json_escape(val_str);
+                    if (val_esc) {
+                        size_t vlen = strlen(val_esc);
+                        memcpy(o, val_esc, vlen);
+                        o += vlen;
+                        free(val_esc);
+                    }
+                    *o++ = '"';
+                } else {
+                    // Fallback to null if conversion failed
+                    memcpy(o, "null", 4);
+                    o += 4;
+                }
+
+                Py_XDECREF(val_str_obj);
+            }
+
+            // Clean up temporary key string object
+            Py_XDECREF(key_str_obj);
+
+            // Clear any exceptions that occurred during conversion
+            if (PyErr_Occurred()) {
+                PyErr_Clear();
+            }
+        }
+        *o++ = '}';
+        *o = '\0';
+        Py_DECREF(items);
+
+        *out_len = (size_t)(o - result);
+        return result;
+    }
+
+    // Handle list of bytes (body chunks)
+    if (PyList_Check(obj)) {
+        Py_ssize_t nchunks = PyList_Size(obj);
+        if (nchunks == 0) {
+            *out_len = 0;
+            return NULL;
+        }
+
+        // First pass: calculate total size
+        size_t total = 0;
+        for (Py_ssize_t i = 0; i < nchunks; i++) {
+            PyObject *chunk = PyList_GetItem(obj, i);
+            if (PyBytes_Check(chunk)) {
+                total += (size_t)PyBytes_Size(chunk);
+            }
+        }
+
+        if (total == 0) {
+            *out_len = 0;
+            return NULL;
+        }
+
+        // Allocate raw buffer
+        char *raw = (char*)malloc(total + 1);
+        if (!raw) {
+            *out_len = 0;
+            return NULL;
+        }
+
+        // Second pass: concatenate
+        char *o = raw;
+        for (Py_ssize_t i = 0; i < nchunks; i++) {
+            PyObject *chunk = PyList_GetItem(obj, i);
+            if (PyBytes_Check(chunk)) {
+                char *data = PyBytes_AsString(chunk);
+                if (!data) {
+                    PyErr_Clear();  // Clear exception and skip this chunk
+                    continue;
+                }
+                Py_ssize_t len = PyBytes_Size(chunk);
+                memcpy(o, data, (size_t)len);
+                o += len;
+            }
+        }
+        *o = '\0';
+
+        // JSON-escape
+        char *escaped = json_escape(raw);
+        free(raw);
+
+        if (escaped) {
+            *out_len = strlen(escaped);
+            return escaped;
+        }
+        *out_len = 0;
+        return NULL;
+    }
+
+    // Handle bytes
+    if (PyBytes_Check(obj)) {
+        char *data = PyBytes_AsString(obj);
+        if (!data) {
+            PyErr_Clear();  // Clear exception
+            *out_len = 0;
+            return NULL;
+        }
+        Py_ssize_t len = PyBytes_Size(obj);
+        char *escaped = json_escape(data);
+        if (escaped) {
+            *out_len = strlen(escaped);
+            return escaped;
+        }
+        *out_len = 0;
+        return NULL;
+    }
+
+    // Handle string
+    if (PyUnicode_Check(obj)) {
+        const char *str = PyUnicode_AsUTF8(obj);
+        if (str) {
+            char *escaped = json_escape(str);
+            if (escaped) {
+                *out_len = strlen(escaped);
+                return escaped;
+            }
+        } else {
+            // Clear exception from PyUnicode_AsUTF8 (invalid UTF-8 string)
+            PyErr_Clear();
+        }
+        *out_len = 0;
+        return NULL;
+    }
+
+    *out_len = 0;
+    return NULL;
+}
+
+// Ultra-fast body capture with GIL-released consolidation
+static PyObject *py_networkhop_with_bodies(PyObject *self, PyObject *args, PyObject *kw) {
+    const char *session_id; Py_ssize_t session_len = 0;
+    int endpoint_id = -1;
+    const char *raw_path = NULL;
+    const char *raw_query_string = NULL; Py_ssize_t raw_query_len = 0;
+    PyObject *req_headers_obj = NULL, *req_body_obj = NULL;
+    PyObject *resp_headers_obj = NULL, *resp_body_obj = NULL;
+
+    static char *kwlist[] = {"session_id", "endpoint_id", "raw_path", "raw_query_string",
+                             "request_headers", "request_body",
+                             "response_headers", "response_body", NULL};
+    if (!PyArg_ParseTupleAndKeywords(args, kw, "s#i|zy#OOOO", kwlist,
+                                     &session_id, &session_len, &endpoint_id,
+                                     &raw_path,
+                                     &raw_query_string, &raw_query_len,
+                                     &req_headers_obj, &req_body_obj,
+                                     &resp_headers_obj, &resp_body_obj)) {
+        return NULL;  // PyArg_ParseTupleAndKeywords sets exception on failure
+    }
+
+    if (!g_running || g_json_prefix == NULL) Py_RETURN_NONE;
+    if (endpoint_id < 0 || endpoint_id >= SFN_ENDPOINT_CAP) Py_RETURN_NONE;
+    if (!g_endpoints[endpoint_id].in_use) Py_RETURN_NONE;
+
+    // Copy session_id with GIL held (small, fast)
+    size_t len = (size_t)session_len;
+    char *sid_copy = (char*)malloc(len + 1);
+    if (!sid_copy) Py_RETURN_NONE;
+    memcpy(sid_copy, session_id, len);
+    sid_copy[len] = 0;
+
+    // JSON-escape raw_path if provided (it's a string)
+    char *route_escaped = NULL;
+    size_t route_len = 0;
+    if (raw_path) {
+        route_escaped = json_escape(raw_path);
+        if (route_escaped) route_len = strlen(route_escaped);
+    }
+
+    // Decode and JSON-escape raw_query_string if provided (it's bytes)
+    char *query_params_escaped = NULL;
+    size_t query_params_len = 0;
+    if (raw_query_string && raw_query_len > 0) {
+        // Decode bytes to string (UTF-8)
+        char *query_str = (char*)malloc((size_t)raw_query_len + 1);
+        if (query_str) {
+            memcpy(query_str, raw_query_string, (size_t)raw_query_len);
+            query_str[raw_query_len] = 0;
+
+            // JSON-escape it
+            query_params_escaped = json_escape(query_str);
+            if (query_params_escaped) query_params_len = strlen(query_params_escaped);
+
+            free(query_str);
+        }
+    }
+
+    // Extract and JSON-escape request/response data (with GIL held)
+    size_t req_headers_len = 0, req_body_len = 0, resp_headers_len = 0, resp_body_len = 0;
+    char *req_headers = extract_and_escape_data(req_headers_obj, &req_headers_len);
+    char *req_body = extract_and_escape_data(req_body_obj, &req_body_len);
+    char *resp_headers = extract_and_escape_data(resp_headers_obj, &resp_headers_len);
+    char *resp_body = extract_and_escape_data(resp_body_obj, &resp_body_len);
+
+    // Check if any Python exceptions occurred during extraction
+    if (PyErr_Occurred()) {
+        // Clean up allocated memory
+        free(sid_copy);
+        free(route_escaped);
+        free(query_params_escaped);
+        free(req_headers);
+        free(req_body);
+        free(resp_headers);
+        free(resp_body);
+        return NULL;  // Propagate the exception
+    }
+
+    // Build message (all data copied, GIL can be released for queue)
+    sfn_msg_t msg = {
+        .type = SFN_MSG_WORK_BODIES,
+        .data.work_bodies = {
+            .session_id = sid_copy,
+            .session_len = len,
+            .endpoint_id = endpoint_id,
+            .route = route_escaped,
+            .route_len = route_len,
+            .query_params = query_params_escaped,
+            .query_params_len = query_params_len,
+            .req_headers = req_headers,
+            .req_headers_len = req_headers_len,
+            .req_body = req_body,
+            .req_body_len = req_body_len,
+            .resp_headers = resp_headers,
+            .resp_headers_len = resp_headers_len,
+            .resp_body = resp_body,
+            .resp_body_len = resp_body_len
+        }
+    };
+
+    // Release GIL for queue operation
+    int pushed = 0;
+    Py_BEGIN_ALLOW_THREADS
+    pushed = ring_try_push(msg);
+    if (!pushed) overflow_push(msg);
+    Py_END_ALLOW_THREADS
+
+    Py_RETURN_NONE;
+}
+
+static PyObject *py_shutdown(PyObject *self, PyObject *args) {
+    if (!g_running) Py_RETURN_NONE;
+    atomic_store(&g_running, 0);
+
+    // Wake ALL threads with broadcast (not signal)
+    pthread_mutex_lock(&g_cv_mtx);
+    pthread_cond_broadcast(&g_cv);
+    pthread_mutex_unlock(&g_cv_mtx);
+
+    // Join all sender threads in thread pool
+    for (int i = 0; i < g_num_sender_threads; i++) {
+        if (g_sender_threads[i]) {
+            pthread_join(g_sender_threads[i], NULL);
+            g_sender_threads[i] = 0;
+        }
+    }
+    g_num_sender_threads = 0;
+
+    // NOTE: g_multi is now per-thread and cleaned by pthread_cleanup_push in sender_main()
+    // No need to clean it here
+
+    // cleanup pool
+    pthread_mutex_lock(&g_pool_mtx);
+    while (g_easy_pool) {
+        pool_node_t *n = g_easy_pool; g_easy_pool = n->next;
+        curl_easy_cleanup(n->easy);
+        free(n);
+    }
+    pthread_mutex_unlock(&g_pool_mtx);
+
+    if (g_share_template) { curl_easy_cleanup(g_share_template); g_share_template = NULL; }
+    if (g_hdrs) { curl_slist_free_all(g_hdrs); g_hdrs = NULL; }
+
+    curl_global_cleanup();
+
+    free(g_url); g_url=NULL;
+    free(g_query_escaped); g_query_escaped=NULL;
+    free(g_json_prefix);   g_json_prefix=NULL;
+    free(g_api_key); g_api_key=NULL;
+    free(g_service_uuid); g_service_uuid=NULL;
+
+    for (int i = 0; i < SFN_ENDPOINT_CAP; ++i) {
+        if (g_endpoints[i].in_use) {
+            free(g_endpoints[i].suffix);
+            g_endpoints[i].suffix = NULL;
+            g_endpoints[i].suffix_len = 0;
+            g_endpoints[i].in_use = 0;
+        }
+    }
+    atomic_store(&g_endpoint_count, 0);
+
+    if (g_ring) {
+        sfn_msg_t msg;
+        while (ring_pop(&msg)) {
+            msg_free(&msg);
+        }
+        free(g_ring); g_ring=NULL;
+    }
+    sfn_node_t *list = overflow_pop_all();
+    while (list) {
+        sfn_node_t *next = list->next;
+        msg_free(&list->msg);
+        free(list);
+        list = next;
+    }
+
+    Py_RETURN_NONE;
+}
+
+static PyMethodDef SFNetworkHopMethods[] = {
+    {"init",                  (PyCFunction)py_init,                  METH_VARARGS | METH_KEYWORDS, "Init networkhop and start sender"},
+    {"register_endpoint",     (PyCFunction)py_register_endpoint,     METH_VARARGS | METH_KEYWORDS, "Register endpoint invariants -> endpoint_id"},
+    {"networkhop",            (PyCFunction)py_networkhop,            METH_VARARGS | METH_KEYWORDS, "Send network hop (generic)"},
+    {"networkhop_fast",       (PyCFunction)py_networkhop_fast,       METH_VARARGS | METH_KEYWORDS, "Send network hop (fast, uses endpoint_id)"},
+    {"networkhop_with_bodies",(PyCFunction)py_networkhop_with_bodies,METH_VARARGS | METH_KEYWORDS, "Send network hop with optional body chunks (GIL-released)"},
+    {"shutdown",              (PyCFunction)py_shutdown,              METH_NOARGS,                   "Shutdown sender and free state"},
+    {NULL, NULL, 0, NULL}
+};
+
+static struct PyModuleDef sfnetworkhopmodule = {
+    PyModuleDef_HEAD_INIT,
+    "_sfnetworkhop",
+    "sf_veritas ultra-fast network hop capture",
+    -1,
+    SFNetworkHopMethods
+};
+
+PyMODINIT_FUNC PyInit__sfnetworkhop(void) {
+    return PyModule_Create(&sfnetworkhopmodule);
+}