fix: node outputs block instead of dropping on a full channel

Every node output used the throwing push(), so a consumer falling behind cost
values rather than time. push_blocking() already existed on Channel and
OutputPort — "wait for the consumer to drain instead of dropping; the producer
just runs slower" — but nothing called it.

A dropped frame does not degrade a downstream result, it silently changes one,
and the consumer has no way to tell it happened. For any pipeline whose output
is a claim about its input, that is corruption rather than degradation.

Safe because sentinels are already handled out-of-band, above this path: only
data blocks, so the EOF token that unwinds the network can always overtake a
stalled data path. That is exactly the hold-and-wait deadlock the push_sentinel
comment warns about, and the reason it is not reachable here.

Measured on a downstream consumer (face pipeline, 77s clip at 5 fps, expected
385 sampled frames):

  before  65 frames written, 320 dropped at one node, 29s
  after   385 frames written, 0 dropped, 17s

Faster, not slower — a dropped frame has already cost its decode, and the
overflow exception cost more. Two consecutive runs now produce byte-identical
output, which they did not before: what got dropped depended on timing, so the
same command could yield different results.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

.gitea/workflows/ci.yaml

@@ -63,7 +63,8 @@ jobs:
     if: ${{ needs.changes.outputs.dockerfile == 'true' }}
     uses: ./.gitea/workflows/docker.yaml
     with:
-      push: ${{ github.event_name != 'pull_request' }}
+      # Explicit string, not a boolean expression (act_runner mangles bools).
+      push: ${{ github.event_name == 'pull_request' && 'false' || 'true' }}
 
   # Runs after docker (if docker ran). A skipped docker job is fine; a failed
   # one blocks this via !failure(). Re-run tests when code OR the image changed.
@@ -72,6 +73,13 @@ jobs:
     if: ${{ !failure() && !cancelled() && (needs.changes.outputs.code == 'true' || needs.changes.outputs.dockerfile == 'true') }}
     uses: ./.gitea/workflows/test.yaml
 
+  # ThreadSanitizer run for the lock-free Channel<T>. Same trigger conditions as
+  # test (code or image changed); runs in parallel with test.
+  tsan:
+    needs: [changes, docker]
+    if: ${{ !failure() && !cancelled() && (needs.changes.outputs.code == 'true' || needs.changes.outputs.dockerfile == 'true') }}
+    uses: ./.gitea/workflows/tsan.yaml
+
   docs:
     needs: [changes, docker]
     if: ${{ !failure() && !cancelled() && github.ref == 'refs/heads/master' && (needs.changes.outputs.docs == 'true' || needs.changes.outputs.dockerfile == 'true') }}

.gitea/workflows/docker.yaml

@@ -4,19 +4,22 @@ name: '🐳 Builder Image'
 # It is called by ci.yaml only when Dockerfile.builder or docs/requirements.txt
 # change. It runs on the host runner (NOT inside the builder container) because
 # it needs the Docker CLI/daemon.
+# Note: `push` is a STRING ("true"/"false"), not a boolean. Gitea's act_runner
+# mangles boolean inputs passed from an expression (they arrive as false), so we
+# pass an explicit string and compare with == 'true' below.
 on:
   workflow_call:
     inputs:
       push:
-        description: 'Push the built image to the registry'
+        description: 'Push the built image to the registry ("true"/"false")'
-        type: boolean
+        type: string
-        default: true
+        default: 'true'
   workflow_dispatch:
     inputs:
       push:
-        description: 'Push the built image to the registry'
+        description: 'Push the built image to the registry ("true"/"false")'
-        type: boolean
+        type: string
-        default: true
+        default: 'true'
 
 jobs:
   build:
@@ -48,7 +51,7 @@ jobs:
             .
 
       - name: Push builder image
-        if: ${{ inputs.push }}
+        if: ${{ inputs.push == 'true' }}
         run: |
           docker push gitea.tourolle.paris/dtourolle/kpnpp-builder:latest
           docker push gitea.tourolle.paris/dtourolle/kpnpp-builder:${{ github.sha }}

.gitea/workflows/tsan.yaml

@@ -0,0 +1,77 @@
+name: '🧵 ThreadSanitizer'
+
+# Reusable workflow: builds the channel stress suite with ThreadSanitizer and
+# runs it. This is the dynamic half of verifying the lock-free SPSC Channel<T>
+# (the static half is the CDSChecker model-check harness in verify/).
+#
+# Triggering and path filtering are owned by ci.yaml (the orchestrator), which
+# calls this only when code changed. workflow_dispatch is kept for manual runs.
+#
+# Runs in the prebuilt builder image (gcc:14), which already ships libtsan — no
+# package installs at job time.
+on:
+  workflow_call:
+  workflow_dispatch:
+
+jobs:
+  tsan:
+    runs-on: linux/amd64
+    container:
+      image: gitea.tourolle.paris/dtourolle/kpnpp-builder:latest
+      # This runner is Docker nested in an unprivileged LXC container, whose
+      # kernel randomizes mmap addresses beyond the range TSan's fixed shadow
+      # mapping expects, so TSan aborts at init with "unexpected memory
+      # mapping". The fix is to disable ASLR per-process with `setarch -R`
+      # (below), which needs the personality(2) syscall that Docker's default
+      # seccomp profile blocks. seccomp=unconfined permits it. Verified on the
+      # runner: setarch -R alone gets EPERM, seccomp alone still aborts, both
+      # together run clean. Scoped to this job, which runs only our own tests.
+      options: --security-opt seccomp=unconfined
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          path: tsan-${{ github.run_id }}
+
+      - name: Cache FetchContent dependencies
+        uses: actions/cache@v3
+        with:
+          path: ~/.cmake/fetchcontent
+          key: cmake-fetchcontent-${{ hashFiles('**/CMakeLists.txt') }}
+          restore-keys: cmake-fetchcontent-
+
+      - name: Configure (TSan)
+        working-directory: tsan-${{ github.run_id }}
+        run: |
+          cmake -S . -B build \
+            -G Ninja \
+            -DCMAKE_BUILD_TYPE=Debug \
+            -DKPN_SANITIZER=thread \
+            -DKPN_BUILD_TESTS=ON \
+            -DKPN_BUILD_EXAMPLES=OFF \
+            -DKPN_BUILD_PYTHON=OFF \
+            -DFETCHCONTENT_BASE_DIR=$HOME/.cmake/fetchcontent
+
+      - name: Build (TSan)
+        working-directory: tsan-${{ github.run_id }}
+        run: cmake --build build --parallel --target kpn_tests kpn_tests_stress
+
+      - name: Run stress suite under TSan
+        working-directory: tsan-${{ github.run_id }}
+        # halt_on_error=1 makes the first detected race fail the job; the report
+        # (with both stacks) is printed to the log. second_deadlock_stack gives
+        # the full picture for lock-order issues.
+        env:
+          TSAN_OPTIONS: "halt_on_error=1 second_deadlock_stack=1"
+        # setarch -R disables ASLR for this process; see the container comment.
+        run: setarch -R ./build/tests/kpn_tests_stress
+
+      - name: Run unit tests under TSan
+        working-directory: tsan-${{ github.run_id }}
+        env:
+          TSAN_OPTIONS: "halt_on_error=1 second_deadlock_stack=1"
+        run: setarch -R ./build/tests/kpn_tests
+
+      - name: Cleanup
+        if: always()
+        run: rm -rf tsan-${{ github.run_id }}

CMakeLists.txt

@@ -10,6 +10,30 @@ option(KPN_BUILD_PYTHON   "Build Python bindings (requires nanobind)" ON)
 option(KPN_BUILD_EXAMPLES "Build examples"                            ON)
 option(KPN_WEB_DEBUG      "Enable web debug UI (cpp-httplib)"         OFF)
 
+# Sanitizer build. Empty = off. Accepts "thread", "address", "undefined",
+# or a combination like "address,undefined". Applied to all kpn targets via
+# the kpn_sanitizer_flags() helper below.
+#
+# The lock-free SPSC Channel<T> (include/kpn/channel.hpp) has hand-reasoned
+# acquire/release ordering; -DKPN_SANITIZER=thread + the channel stress test
+# (tests/test_channel_stress.cpp) is the dynamic half of verifying it. The
+# static half is the CDSChecker model-check harness (see verify/).
+set(KPN_SANITIZER "" CACHE STRING
+    "Build with sanitizer: thread | address | undefined | <combo> (empty = off)")
+
+# Translate KPN_SANITIZER into compile/link flags. No-op when empty.
+function(kpn_sanitizer_flags out_var)
+    if(KPN_SANITIZER)
+        set(${out_var}
+            -fsanitize=${KPN_SANITIZER}
+            -fno-omit-frame-pointer
+            -g
+            PARENT_SCOPE)
+    else()
+        set(${out_var} "" PARENT_SCOPE)
+    endif()
+endfunction()
+
 # ── Core library (header-only) ────────────────────────────────────────────────
 add_library(kpn INTERFACE)
 target_include_directories(kpn INTERFACE

Dockerfile.builder

@@ -1,4 +1,4 @@
-# KPN++ Builder Image  (CI: pipeline trigger)
+# KPN++ Builder Image  (CI: pipeline trigger v2)
 # Pre-built image with GCC, CMake, Ninja, and Python dev headers for building and testing KPN++
 # Build: docker build -f Dockerfile.builder -t gitea.tourolle.paris/dtourolle/kpnpp-builder:latest .
 # Push:  docker push gitea.tourolle.paris/dtourolle/kpnpp-builder:latest

README.md

@@ -2,6 +2,8 @@
 
 A C++20 Kahn Process Network (KPN) library. Each node wraps a function and runs in its own thread, communicating with downstream nodes via bounded FIFO channels. Includes Python bindings via nanobind.
 
+📖 **[Documentation](https://pages.tourolle.paris/dtourolle/kpn/)**
+
 ---
 
 ## Requirements
@@ -398,8 +400,8 @@ Violating the second rule deadlocks.
 | `04_storage_policy` | `channel_storage_policy` default and specialisation |
 | `05_error_handling` | `ChannelOverflowError`, `ErrorHandler` |
 | `06_watchdog` | Watchdog interval, stall detection |
-| `07_python_network` | PyNetwork, pure Python node *(pending)* |
+| `07_python_network` | PyNetwork with a pure-Python node between a C++ source and sink |
-| `08_python_subport` | `net.read`, `net.write`, sub-port tap *(pending)* |
+| `08_python_subport` | Drive a Python node from Python via `net.write`/`net.read` sub-port taps |
 | `09_opencv_cellshade` | Real-time cell-shading on webcam/pattern; requires OpenCV ≥ 4 |
 
 Run the cell-shading example:

README.md.in

@@ -2,6 +2,8 @@
 
 A C++20 Kahn Process Network (KPN) library. Each node wraps a function and runs in its own thread, communicating with downstream nodes via bounded FIFO channels. Includes Python bindings via nanobind.
 
+📖 **[Documentation](https://pages.tourolle.paris/dtourolle/kpn/)**
+
 ---
 
 ## Requirements
@@ -220,8 +222,8 @@ Violating the second rule deadlocks.
 | `04_storage_policy` | `channel_storage_policy` default and specialisation |
 | `05_error_handling` | `ChannelOverflowError`, `ErrorHandler` |
 | `06_watchdog` | Watchdog interval, stall detection |
-| `07_python_network` | PyNetwork, pure Python node *(pending)* |
+| `07_python_network` | PyNetwork with a pure-Python node between a C++ source and sink |
-| `08_python_subport` | `net.read`, `net.write`, sub-port tap *(pending)* |
+| `08_python_subport` | Drive a Python node from Python via `net.write`/`net.read` sub-port taps |
 | `09_opencv_cellshade` | Real-time cell-shading on webcam/pattern; requires OpenCV ≥ 4 |
 
 Run the cell-shading example:

examples/07_python_network/example.py

@@ -30,3 +30,8 @@ net.build()
 net.start()
 time.sleep(0.1)
 net.stop()
+
+# Drop the network deterministically: it holds the Python callable, which forms
+# a reference cycle via globals(). Deleting the global breaks it so the network
+# is reclaimed now rather than lingering to interpreter shutdown.
+del net

examples/08_python_subport/example.py

@@ -1,38 +1,55 @@
 """
-08_python_subport — tap a C++ node's output from Python using net.read().
+08_python_subport — drive a *Python* node from Python via write()/read() taps.
 
 Graph:
-  [ProduceNode] --int--> [DoubleItNode] --int--> (tapped by net.read())
+  (fed by net.write()) --int--> [py_triple] --int--> (tapped by net.read())
 
-The sink is Python: instead of connecting a PrintItNode, we call net.read()
+Unlike 07, there is no C++ source or sink here: the only node in the network is
-to pull values out of DoubleItNode's output directly into Python.
+a pure-Python function, py_triple. Python plays *both* the producer and the
-We also demonstrate net.write() by injecting a value into DoubleItNode's input.
+consumer by using the subport taps:
+
+  * net.write("py", 0, v)  injects v into py_triple's input   (Python -> network)
+  * net.read("py", 0)      pulls py_triple's output back out   (network -> Python)
+
+This closes the loop the old version left as a "#todo": a value flows from
+Python, through a Python node running inside the network, and back to Python.
 """
 
 import sys
-import time
+sys.path.insert(0, "build/python")  # for `python examples/.../example.py` from repo root
-import threading
-sys.path.insert(0, "build/python")
 
 import kpn_python as kpn
 
+
+def py_triple(x: int) -> int:
+    return x * 3
+
+
 net = kpn.Network()
 
-net.add("src", kpn.make_produce())
+# The whole network is a single Python node with a tapped input and output.
-net.add("dbl", kpn.make_double_it())
+net.add_node("py", py_triple, inputs=["int"], outputs=["int"])
 
-net.connect("src", 0, "dbl", 0)
 net.build()
 net.start()
 
-# Collect a few values from DoubleItNode's output via Python tap
+# Push values in from Python and read the Python node's results back out.
+inputs = [1, 2, 7, 10, 100]
 results = []
-for _ in range(5):
+for v in inputs:
-    val = net.read("dbl", 0)
+    net.write("py", 0, v)          # Python -> py_triple input
-    results.append(val)
+    results.append(net.read("py", 0))  # py_triple output -> Python
 
 net.stop()
 
-print("values read from C++ DoubleItNode output:", results)
+print("inputs written from Python: ", inputs)
-assert all(v == 84 for v in results), f"expected all 84, got {results}"
+print("outputs read from py_triple:", results)
-print("all correct (42 * 2 = 84)")
+
+expected = [v * 3 for v in inputs]
+assert results == expected, f"expected {expected}, got {results}"
+print("all correct (x * 3 computed by a Python node inside the network)")
+
+# Drop the network deterministically. The network holds the Python callable,
+# which (via globals) forms a reference cycle; deleting the global breaks it so
+# the network is reclaimed promptly instead of lingering to interpreter exit.
+del net

examples/CMakeLists.txt

@@ -13,6 +13,25 @@ function(kpn_example name)
     )
 endfunction()
 
+# Register a Python example script as a CTest smoke test. Runs the script with
+# PYTHONPATH pointing at the freshly-built kpn_python module, so it does not
+# depend on the caller's working directory or a hard-coded "build/python" path.
+function(kpn_python_example name)
+    if(NOT KPN_BUILD_PYTHON)
+        return()
+    endif()
+    add_test(
+        NAME    example_${name}
+        COMMAND ${CMAKE_COMMAND} -E env
+                "PYTHONPATH=$<TARGET_FILE_DIR:kpn_python>"
+                ${Python_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/${name}/example.py
+    )
+    set_tests_properties(example_${name} PROPERTIES
+        TIMEOUT 15
+        LABELS  examples
+    )
+endfunction()
+
 kpn_example(01_hello_pipeline)
 kpn_example(02_named_ports)
 kpn_example(03_multi_output)
@@ -31,7 +50,9 @@ if(KPN_WEB_DEBUG)
     target_link_libraries(14_debug_hub PRIVATE kpn)
     kpn_target_enable_web_debug(14_debug_hub)
 endif()
-# 07 and 08 are Python scripts — no compiled target needed.
+# 07 and 08 are Python scripts — no compiled target, but run as smoke tests.
+kpn_python_example(07_python_network)
+kpn_python_example(08_python_subport)
 
 # 09 requires OpenCV — only build if found
 find_package(OpenCV QUIET COMPONENTS core imgproc highgui videoio)

include/kpn/channel.hpp

@@ -136,6 +136,35 @@ public:
             push_callback_();
     }
 
+    // Lossless push with BACKPRESSURE: if the ring is full, wait for the consumer to
+    // drain instead of dropping (the throwing push()) — the producer just runs slower.
+    // Use when every value must be delivered (e.g. replaying a dump for scoring, where
+    // a dropped frame silently corrupts the result). SPSC: only the sole producer may
+    // call it. Returns false if the channel was disabled while waiting.
+    bool push_blocking(T value) {
+        for (;;) {
+            if (!accepting_.load(std::memory_order_acquire)) {
+                stats_.record_drop();
+                return false;
+            }
+            const std::size_t t = tail_.load(std::memory_order_relaxed);
+            const std::size_t h = head_.load(std::memory_order_acquire);
+            if (t - h < capacity_) {          // space available → normal push
+                const std::size_t data_bytes = ChannelDataSize<T>::bytes(value);
+                const bool was_empty = (t == h);
+                buf_[t & ring_mask_] = make_storage(std::move(value));
+                tail_.store(t + 1, std::memory_order_release);
+                stats_.record_push(t - h + 1, data_bytes);
+                wake_.fetch_add(1, std::memory_order_release);
+                wake_.notify_one();
+                if (was_empty && push_callback_) push_callback_();
+                return true;
+            }
+            // full: yield briefly and retry (consumer will drain)
+            std::this_thread::sleep_for(std::chrono::microseconds(50));
+        }
+    }
+
     // Lossless, non-blocking delivery for a must-deliver control token (EOF).
     //
     // A sentinel is stored out-of-band — in a dedicated slot that does NOT
@@ -180,7 +209,16 @@ public:
             if (h == t) {
                 // Ring drained — deliver any pending out-of-band sentinel (EOF)
                 // now, so it always arrives after the data pushed before it.
-                { T s; if (take_sentinel(s)) return s; }
+                //
+                // Re-confirm emptiness against a fresh tail_ first: the snapshot
+                // at the top of the loop may be stale (the producer can push more
+                // values *and* the sentinel in the window since), and the sentinel
+                // must never jump ahead of ring values pushed before it. The spin
+                // and post-spin takes below already reload tail_ on the line above
+                // them; this is the one take that used the loop-top snapshot.
+                if (h == tail_.load(std::memory_order_acquire)) {
+                    T s; if (take_sentinel(s)) return s;
+                }
 
                 if (!accepting_.load(std::memory_order_acquire))
                     throw ChannelClosedError{};

include/kpn/pool_node.hpp

@@ -400,12 +400,15 @@ private:
             ch->push_sentinel(std::move(val));
             return;
         }
-        try {
+        // Backpressure, not loss. A full downstream channel means the consumer
-            ch->push(std::move(val));
+        // is behind, and the correct response is for this producer to run
-        } catch (const ChannelOverflowError&) {
+        // slower — not to discard a value. A dropped frame does not degrade a
-            throw ChannelOverflowError(ch->capacity(),
+        // result, it silently changes one, and the caller has no way to tell.
-                "pool node '" + name_ + "' " + output_port_label<I>());
+        //
-        }
+        // Safe here because sentinels are handled above, out-of-band: this
+        // blocks only on data, so the EOF token that unwinds the network can
+        // always overtake a stalled data path.
+        ch->push_blocking(std::move(val));
     }
 
     template<std::size_t I>
@@ -706,12 +709,8 @@ private:
             ch->push_sentinel(std::move(val));
             return;
         }
-        try {
+        // See the note on the typed overload above: block rather than drop.
-            ch->push(std::move(val));
+        ch->push_blocking(std::move(val));
-        } catch (const ChannelOverflowError&) {
-            throw ChannelOverflowError(ch->capacity(),
-                "pool node '" + name_ + "'");
-        }
     }
 
     Obj&                                     obj_;

include/kpn/python/auto_bind.hpp

@@ -247,7 +247,7 @@ void bind_network(nb::module_& m) {
 
     nb::class_<IVariantNode<Variant>>(m, "INode");
 
-    nb::class_<Net>(m, "Network")
+    nb::class_<Net>(m, "Network", nb::type_slots(network_type_slots<Variant>()))
         .def("__init__", [](Net* self) {
             new (self) Net();
             register_all_converters<Registry>(*self);

include/kpn/python/bindings.hpp

@@ -35,6 +35,16 @@ public:
     using VNode    = IVariantNode<Variant>;
     using VChannel = IVariantChannel<Variant>;
 
+    // ── GC support ────────────────────────────────────────────────────────────
+    // Visit every Python object this network transitively holds (currently the
+    // callable of each PyNode). Used by the Network type's tp_traverse slot so
+    // Python's cyclic GC can discover instance → callable → globals() cycles.
+    // Defined out-of-line below, once PyNode is a complete type.
+    template<typename Fn>
+    void visit_python_objects(Fn&& visit) const;
+    // Drop all Python references held by nodes, breaking any cycle (tp_clear).
+    void clear_python_objects();
+
     // ── Builder API ───────────────────────────────────────────────────────────
 
     void add(std::string name, std::shared_ptr<VNode> node) {
@@ -207,6 +217,20 @@ public:
         return it->second;
     }
 
+    // Raw node handle by name — lets a binding dynamic_cast to a concrete wrapper
+    // type and call its functor's runtime setters (persistent-pipeline reuse).
+    VNode* node_ptr(const std::string& name) { return &node_at(name); }
+
+    // Per-node timing snapshot for profiling where a replay spends its time.
+    std::map<std::string, double> node_stats(const std::string& name) {
+        auto& n = node_at(name);
+        NodeSnapshot s = n.node_snapshot(name, 0.0);
+        return {{"frames", double(s.frames_processed)},
+                {"exec_ms", s.ema_exec_ms}, {"max_ms", s.max_exec_ms},
+                {"blocked_ms", s.total_blocked_ms}, {"fps", s.throughput_fps},
+                {"cpu_ms", s.total_cpu_ms}, {"cpu_util_pct", s.cpu_util_pct}};
+    }
+
 private:
     VNode& node_at(const std::string& name) {
         auto it = nodes_.find(name);
@@ -367,6 +391,14 @@ public:
         out_channels_[i] = std::move(ch);
     }
 
+    // ── GC support (tp_traverse / tp_clear on the owning Network) ──────────────
+    // The node holds a Python callable, which typically forms an
+    // instance → callable → globals() → instance cycle. Expose the callable so
+    // the Network's GC slots can traverse and clear it. See bindings.hpp's
+    // network_tp_traverse/network_tp_clear.
+    const nb::object& python_callable() const { return callable_; }
+    void clear_python_callable() { callable_ = nb::object(); }
+
 private:
     void run_loop() {
         while (!stop_flag_.load(std::memory_order_relaxed)) {
@@ -404,13 +436,17 @@ private:
 
                 for (std::size_t i = 0; i < out_channels_.size(); ++i) {
                     if (out_channels_[i])
-                        out_channels_[i]->push(std::move(outputs[i]));
+                        // Lossless: wait for space rather than drop. A dropped frame
+                        // silently corrupts a replay's score; backpressure just slows
+                        // the producer. (Was push() + "drop on overflow".)
+                        out_channels_[i]->push_blocking(std::move(outputs[i]));
                 }
 
             } catch (const ChannelClosedError&) {
                 break;
             } catch (const ChannelOverflowError&) {
-                // drop and continue
+                // no longer reachable with push_blocking, kept for safety
+                break;
             }
         }
     }
@@ -439,6 +475,60 @@ private:
     NodeStats          stats_;
 };
 
+// ── PyNetwork GC helpers (defined here: PyNode is now complete) ────────────────
+
+template<typename Variant>
+template<typename Fn>
+void PyNetwork<Variant>::visit_python_objects(Fn&& visit) const {
+    for (const auto& [name, node] : nodes_)
+        if (auto* py = dynamic_cast<const PyNode<Variant>*>(node.get()))
+            visit(py->python_callable());
+}
+
+template<typename Variant>
+void PyNetwork<Variant>::clear_python_objects() {
+    for (auto& [name, node] : nodes_)
+        if (auto* py = dynamic_cast<PyNode<Variant>*>(node.get()))
+            py->clear_python_callable();
+}
+
+// ── GC type slots for the Network binding ─────────────────────────────────────
+// The Network holds Python callables (via PyNode), forming uncollectable
+// instance → callable → globals() → instance cycles at interpreter shutdown.
+// These slots let Python's cyclic collector traverse and break them, silencing
+// nanobind's leak warnings. See the nanobind "Reference leaks" documentation.
+
+template<typename Variant>
+int network_tp_traverse(PyObject* self, visitproc visit, void* arg) {
+    Py_VISIT(Py_TYPE(self));
+    if (!nb::inst_ready(self))
+        return 0;
+    auto* net = nb::inst_ptr<PyNetwork<Variant>>(self);
+    int rv = 0;
+    net->visit_python_objects([&](const nb::object& obj) {
+        if (rv == 0 && obj.is_valid())
+            rv = visit(obj.ptr(), arg);
+    });
+    return rv;
+}
+
+template<typename Variant>
+int network_tp_clear(PyObject* self) {
+    auto* net = nb::inst_ptr<PyNetwork<Variant>>(self);
+    net->clear_python_objects();
+    return 0;
+}
+
+template<typename Variant>
+PyType_Slot* network_type_slots() {
+    static PyType_Slot slots[] = {
+        { Py_tp_traverse, reinterpret_cast<void*>(&network_tp_traverse<Variant>) },
+        { Py_tp_clear,    reinterpret_cast<void*>(&network_tp_clear<Variant>) },
+        { 0, nullptr }
+    };
+    return slots;
+}
+
 // ── register_py_network (legacy helper) ───────────────────────────────────────
 // Registers PyNetwork<Variant> with the given nanobind module.
 // Prefer bind_network<Registry> from auto_bind.hpp for new code.
@@ -447,7 +537,7 @@ template<typename Variant>
 void register_py_network(nb::module_& m, const char* class_name = "Network") {
     using Net = PyNetwork<Variant>;
 
-    nb::class_<Net>(m, class_name)
+    nb::class_<Net>(m, class_name, nb::type_slots(network_type_slots<Variant>()))
         .def(nb::init<>())
         .def("connect", &Net::connect,
              nb::arg("src"), nb::arg("out_idx"),
@@ -458,7 +548,8 @@ void register_py_network(nb::module_& m, const char* class_name = "Network") {
         .def("read",   &Net::read,
              nb::arg("node"), nb::arg("out_idx") = std::size_t(0))
         .def("write",  &Net::write,
-             nb::arg("node"), nb::arg("in_idx"), nb::arg("value"));
+             nb::arg("node"), nb::arg("in_idx"), nb::arg("value"))
+        .def("node_stats", &Net::node_stats, nb::arg("node"));
 }
 
 } // namespace kpn::python

include/kpn/python/object_variant_node.hpp

@@ -0,0 +1,149 @@
+#pragma once
+// ObjectVariantNodeWrapper — variant-node adapter for *stateful* functors.
+//
+// VariantNodeWrapper (variant_node.hpp) wraps Node<Func,...>, where Func is a
+// default-constructible NTTP callable. That doesn't fit nodes whose functor must
+// be constructed with runtime state (a Config, a loaded gallery, etc.) — those use
+// ObjectNode<Obj>, which takes `Obj& obj` at construction.
+//
+// This wrapper owns an Obj instance and exposes the same IVariantNode surface so a
+// stateful C++ node can live inside a PyNetwork. Build one via a factory that
+// constructs the functor from Python-supplied config, e.g.:
+//
+//   auto n = std::make_shared<ObjectVariantNodeWrapper<
+//                IdentityMatcherFunc, Variant, in<"tracked">, out<"matched">>>(
+//                fifo_cap, gallery, cfg);            // Obj ctor args forwarded
+//   net.add("identity_matcher", n);
+//
+// The wrapper mirrors VariantNodeWrapper's channel plumbing exactly; only the
+// underlying node type (PoolObjectNode, holding Obj&) differs.
+
+#include "../channel.hpp"
+#include "../node.hpp"
+#include "../variant_node.hpp"
+
+#include <memory>
+#include <stdexcept>
+#include <string>
+#include <tuple>
+#include <typeindex>
+#include <utility>
+#include <vector>
+
+namespace kpn {
+
+template<typename Obj, typename Variant,
+         typename InputTag  = in<>,
+         typename OutputTag = out<>>
+class ObjectVariantNodeWrapper;
+
+template<typename Obj, typename Variant,
+         fixed_string... InNames, fixed_string... OutNames>
+class ObjectVariantNodeWrapper<Obj, Variant, in<InNames...>, out<OutNames...>>
+    : public IVariantNode<Variant>
+{
+    using NodeT = ObjectNode<Obj, in<InNames...>, out<OutNames...>>;
+
+public:
+    using args_tuple   = typename NodeT::args_tuple;
+    using return_tuple = typename NodeT::return_tuple;
+
+    static constexpr std::size_t n_in  = NodeT::input_count;
+    static constexpr std::size_t n_out = NodeT::output_count;
+
+    // Owns the functor; forwards remaining args to Obj's constructor.
+    template<typename... ObjArgs>
+    explicit ObjectVariantNodeWrapper(std::size_t fifo_capacity, ObjArgs&&... obj_args)
+        : obj_(std::forward<ObjArgs>(obj_args)...)
+        , node_(obj_, fifo_capacity)
+        , in_channels_(n_in)
+        , out_channels_(n_out)
+        , out_type_indices_(n_out, std::type_index(typeid(void)))
+    {
+        init_inputs(std::make_index_sequence<n_in>{}, fifo_capacity);
+        init_out_types(std::make_index_sequence<n_out>{});
+    }
+
+    // Access the owned functor so callers can invoke its runtime setters (e.g. to
+    // change a threshold on a persistent pipeline without rebuilding the node).
+    Obj& functor() { return obj_; }
+
+    // ── INode ─────────────────────────────────────────────────────────────────
+    void start() override { node_.start(); }
+    void stop()  override { node_.stop(); }
+    bool running() const override { return node_.running(); }
+    const NodeStats& stats() const override { return node_.stats(); }
+    void set_name(std::string name) override { node_.set_name(std::move(name)); }
+    NodeSnapshot node_snapshot(const std::string& name, double elapsed_s) const override {
+        return node_.node_snapshot(name, elapsed_s);
+    }
+
+    // ── IVariantNode ──────────────────────────────────────────────────────────
+    std::size_t input_count()  const override { return n_in; }
+    std::size_t output_count() const override { return n_out; }
+
+    std::type_index input_type(std::size_t i) const override {
+        return in_channels_[i]->type_index();
+    }
+    std::type_index output_type(std::size_t i) const override {
+        return out_type_indices_[i];
+    }
+
+    std::shared_ptr<IVariantChannel<Variant>> input_channel(std::size_t i) override {
+        return in_channels_[i];
+    }
+
+    void set_output_channel(std::size_t i,
+                            std::shared_ptr<IVariantChannel<Variant>> ch) override {
+        set_output_impl(i, std::move(ch), std::make_index_sequence<n_out>{});
+    }
+
+private:
+    template<std::size_t... Is>
+    void init_inputs(std::index_sequence<Is...>, std::size_t cap) {
+        ((init_one_input<Is>(cap)), ...);
+    }
+
+    template<std::size_t I>
+    void init_one_input(std::size_t cap) {
+        using T = std::tuple_element_t<I, args_tuple>;
+        auto shared_ch = std::make_shared<Channel<T>>(cap);
+        node_.template set_input_channel<I>(shared_ch);
+        in_channels_[I] = std::make_shared<VariantChannel<T, Variant>>(std::move(shared_ch));
+    }
+
+    template<std::size_t... Is>
+    void init_out_types(std::index_sequence<Is...>) {
+        ((out_type_indices_[Is] =
+              std::type_index(typeid(std::tuple_element_t<Is, return_tuple>))), ...);
+    }
+
+    template<std::size_t... Is>
+    void set_output_impl(std::size_t port,
+                         std::shared_ptr<IVariantChannel<Variant>> ch,
+                         std::index_sequence<Is...>) {
+        bool matched = false;
+        ((Is == port && (set_output_at<Is>(std::move(ch)), matched = true)), ...);
+        if (!matched)
+            throw std::out_of_range("set_output_channel: port index out of range");
+    }
+
+    template<std::size_t I>
+    void set_output_at(std::shared_ptr<IVariantChannel<Variant>> ch) {
+        using T = std::tuple_element_t<I, return_tuple>;
+        auto* typed = dynamic_cast<VariantChannel<T, Variant>*>(ch.get());
+        if (!typed)
+            throw std::runtime_error(
+                "set_output_channel: type mismatch at output port " + std::to_string(I));
+        node_.template set_output_channel<I>(typed->raw_ptr());
+        out_channels_[I] = std::move(ch);
+    }
+
+    Obj   obj_;    // owned; node_ holds Obj& — declaration order keeps obj_ alive first
+    NodeT node_;
+    std::vector<std::shared_ptr<IVariantChannel<Variant>>> in_channels_;
+    std::vector<std::shared_ptr<IVariantChannel<Variant>>> out_channels_;
+    std::vector<std::type_index>                           out_type_indices_;
+};
+
+} // namespace kpn

include/kpn/variant_node.hpp

@@ -55,6 +55,8 @@ class IVariantChannel {
 public:
     virtual ~IVariantChannel() = default;
     virtual void            push(Variant v) = 0;
+    // Lossless push with backpressure (waits instead of dropping when full).
+    virtual void            push_blocking(Variant v) = 0;
     virtual Variant         pop()           = 0;
     virtual std::type_index type_index()    const = 0;
     virtual std::string     type_name()     const = 0;
@@ -76,6 +78,9 @@ public:
     void push(Variant v) override {
         channel_->push(std::get<T>(std::move(v)));
     }
+    void push_blocking(Variant v) override {
+        channel_->push_blocking(std::get<T>(std::move(v)));
+    }
     Variant pop() override {
         return Variant{ channel_->pop() };
     }

tests/CMakeLists.txt

@@ -43,6 +43,35 @@ target_link_libraries(kpn_tests PRIVATE
     GTest::gtest
 )
 
+# ── Channel stress suite (separate executable) ────────────────────────────────
+# Contended SPSC tests for the lock-free Channel<T>. Kept out of kpn_tests
+# because each case runs many reps / tens of thousands of items and is slow.
+# Most valuable under -DKPN_SANITIZER=thread, but correct (and run) without it.
+add_executable(kpn_tests_stress test_channel_stress.cpp)
+target_link_libraries(kpn_tests_stress PRIVATE kpn Catch2::Catch2WithMain)
+
+# ── Sanitizer flags ───────────────────────────────────────────────────────────
+# kpn_sanitizer_flags() is defined in the top-level CMakeLists and is a no-op
+# unless -DKPN_SANITIZER=... is set. Sanitizer must be on both compile and link.
+kpn_sanitizer_flags(_kpn_san)
+if(_kpn_san)
+    foreach(_t kpn_tests kpn_tests_stress)
+        target_compile_options(${_t} PRIVATE ${_kpn_san})
+        target_link_options(${_t}    PRIVATE ${_kpn_san})
+    endforeach()
+endif()
+
 include(CTest)
 include(Catch)
-catch_discover_tests(kpn_tests)
+
+# DISCOVERY_MODE PRE_TEST defers test enumeration to `ctest` run time. The
+# default (POST_BUILD) runs each test binary during the build to list its
+# cases — which fails a sanitizer build: a TSan/ASan binary needs a fixed
+# address-space layout and aborts on startup ("unexpected memory mapping")
+# under the container's ASLR, breaking the build before any test runs. The
+# tsan.yaml job invokes the binaries directly (not via ctest), so deferring
+# discovery costs nothing there and keeps `ctest` working for normal builds.
+catch_discover_tests(kpn_tests DISCOVERY_MODE PRE_TEST)
+# Register the stress suite under its own label so CI can run / time it
+# separately from the fast unit tests.
+catch_discover_tests(kpn_tests_stress DISCOVERY_MODE PRE_TEST PROPERTIES LABELS "stress")

tests/test_channel_stress.cpp

@@ -0,0 +1,281 @@
+// Contended stress tests for the lock-free SPSC Channel<T>.
+//
+// The other channel tests (test_channel.cpp) are single-threaded or use a
+// single 20 ms sleep to order two threads — they never actually contend on the
+// ring, so they exercise neither the memory-ordering pairing nor the
+// spin/futex/lost-wakeup logic in pop().
+//
+// These tests are written to be run under ThreadSanitizer:
+//
+//   cmake -B build -DKPN_SANITIZER=thread -DKPN_BUILD_EXAMPLES=OFF -DKPN_BUILD_PYTHON=OFF
+//   cmake --build build --target kpn_tests_tsan
+//   ./build/tests/kpn_tests_tsan
+//
+// They are also valid (and meaningful) without a sanitizer: the value/sequence
+// assertions catch lost or duplicated items regardless of build flags. TSan
+// adds detection of the underlying data race even on runs where the race did
+// not corrupt observable state.
+//
+// Channel<T> is SPSC: exactly one producer thread and one consumer thread per
+// channel. Every scenario below honours that contract.
+
+#include <catch2/catch_test_macros.hpp>
+#include <atomic>
+#include <chrono>
+#include <kpn/channel.hpp>
+#include <thread>
+#include <vector>
+
+using namespace kpn;
+using namespace std::chrono_literals;
+
+namespace {
+
+// Repeat each scenario enough times that rare interleavings (spin window just
+// missing / just catching the next push, disable landing inside the futex
+// wait) actually occur across a run. Kept modest so a TSan run stays minutes,
+// not hours.
+constexpr int kReps = 200;
+
+} // namespace
+
+TEST_CASE("SPSC: every pushed item is popped exactly once, in order", "[channel][stress]") {
+    // Small capacity forces frequent full/empty transitions, so both the
+    // producer's overflow-retry and the consumer's spin->futex path are hit
+    // many times. The producer retries on overflow rather than dropping, so
+    // the consumer must observe a strictly contiguous 0..N-1 sequence.
+    constexpr int N = 50'000;
+    Channel<int> ch(/*capacity=*/4, /*spin_count=*/16);
+
+    std::thread producer([&] {
+        for (int i = 0; i < N; ++i) {
+            for (;;) {
+                try { ch.push(i); break; }
+                catch (const ChannelOverflowError&) { std::this_thread::yield(); }
+            }
+        }
+    });
+
+    int expected = 0;
+    bool in_order = true;
+    for (int i = 0; i < N; ++i) {
+        int v = ch.pop();
+        if (v != expected) in_order = false;
+        ++expected;
+    }
+    producer.join();
+
+    REQUIRE(in_order);
+    REQUIRE(expected == N);
+    REQUIRE(ch.size() == 0);
+}
+
+TEST_CASE("SPSC: tight empty<->non-empty transitions exercise spin/futex boundary",
+          "[channel][stress]") {
+    // spin_count=0 forces every empty pop() straight into atomic::wait, so this
+    // hammers the lost-wakeup guard (snapshot wake_, re-check tail_, then wait).
+    // The producer pushes one item then waits to go empty again, maximising the
+    // number of empty->non-empty edges relative to item count.
+    constexpr int N = 20'000;
+    Channel<int> ch(/*capacity=*/2, /*spin_count=*/0);
+
+    std::thread producer([&] {
+        for (int i = 0; i < N; ++i) {
+            for (;;) {
+                try { ch.push(i); break; }
+                catch (const ChannelOverflowError&) { std::this_thread::yield(); }
+            }
+        }
+    });
+
+    long sum = 0;
+    for (int i = 0; i < N; ++i) sum += ch.pop();
+    producer.join();
+
+    // Sum of 0..N-1 — detects any lost or duplicated item.
+    REQUIRE(sum == static_cast<long>(N) * (N - 1) / 2);
+}
+
+TEST_CASE("SPSC: disable() while consumer is blocked in pop() unblocks cleanly",
+          "[channel][stress]") {
+    // The data race of record: consumer blocked in pop() (spinning or parked in
+    // the futex) while the owner thread calls disable(). pop() must observe the
+    // close and throw ChannelClosedError — it must not hang and must not read
+    // past the ring. Repeated so disable() lands at many points in pop()'s loop.
+    for (int rep = 0; rep < kReps; ++rep) {
+        Channel<int> ch(/*capacity=*/4, /*spin_count=*/8);
+        std::atomic<bool> threw{false};
+        std::atomic<bool> finished{false};
+
+        std::thread consumer([&] {
+            try {
+                ch.pop();              // empty channel: will block
+            } catch (const ChannelClosedError&) {
+                threw.store(true, std::memory_order_relaxed);
+            }
+            finished.store(true, std::memory_order_relaxed);
+        });
+
+        // Give the consumer a chance to reach the wait, then close.
+        std::this_thread::sleep_for(50us);
+        ch.disable();
+
+        consumer.join();
+        REQUIRE(finished.load());
+        REQUIRE(threw.load());
+    }
+}
+
+TEST_CASE("SPSC: producer racing a disable() never throws and never hangs",
+          "[channel][stress]") {
+    // Mirror of the above from the producer side: push() racing disable() must
+    // either enqueue or silently drop, never throw ChannelClosedError and never
+    // wedge. Overflow is still a legal outcome (full accepting channel) and is
+    // tolerated here.
+    for (int rep = 0; rep < kReps; ++rep) {
+        Channel<int> ch(/*capacity=*/8, /*spin_count=*/8);
+        std::atomic<bool> bad{false};
+
+        std::thread producer([&] {
+            for (int i = 0; i < 1000; ++i) {
+                try { ch.push(i); }
+                catch (const ChannelOverflowError&) { /* legal: full */ }
+                catch (...) { bad.store(true, std::memory_order_relaxed); break; }
+            }
+        });
+
+        std::this_thread::sleep_for(20us);
+        ch.disable();           // owner closes mid-stream
+        producer.join();
+
+        REQUIRE_FALSE(bad.load());
+    }
+}
+
+TEST_CASE("SPSC: push_callback fires on each empty->non-empty transition",
+          "[channel][stress]") {
+    // The empty->non-empty callback ([channel.hpp] was_empty branch) is read by
+    // the consumer-side notification path. Run it under contention to make sure
+    // the was_empty detection isn't torn by a concurrent pop().
+    Channel<int> ch(/*capacity=*/4, /*spin_count=*/4);
+    std::atomic<int> callbacks{0};
+    ch.set_push_callback([&] { callbacks.fetch_add(1, std::memory_order_relaxed); });
+
+    constexpr int N = 10'000;
+    std::thread producer([&] {
+        for (int i = 0; i < N; ++i) {
+            for (;;) {
+                try { ch.push(i); break; }
+                catch (const ChannelOverflowError&) { std::this_thread::yield(); }
+            }
+        }
+    });
+
+    for (int i = 0; i < N; ++i) (void)ch.pop();
+    producer.join();
+
+    // At least one transition, at most one per item; mainly we assert the run
+    // completed without TSan flagging a race on push_callback_/was_empty.
+    REQUIRE(callbacks.load() >= 1);
+    REQUIRE(callbacks.load() <= N);
+}
+
+// Ordering contract of the out-of-band sentinel under contention.
+//
+// push_sentinel() publishes has_eof_ (release) after the producer's N ring
+// pushes; a consumer that observes has_eof_ (acquire) therefore also observes
+// every value pushed before it. Both pop() and try_pop_now() only surface the
+// sentinel once the ring is *freshly* observed empty, so the sentinel is the
+// strictly last item received — it never jumps ahead of a ring value pushed
+// before it. These tests treat the sentinel as a hard "last message" barrier
+// (the consumer stops draining the moment it sees it) and assert that all N
+// values arrived, in a contiguous 0..N-1 sequence, before it.
+//
+// Regression guard: an earlier version of pop() checked emptiness against a
+// stale tail_ snapshot from the top of its loop, so under load the sentinel
+// could surface with a few real values still queued — breaking in_order /
+// values==N here. Under TSan these also cover the has_eof_/eof_value_
+// acquire/release handshake and the spin/futex wakeup on push_sentinel().
+
+TEST_CASE("SPSC: sentinel is strictly last, after every value (blocking pop)",
+          "[channel][stress]") {
+    constexpr int N        = 20'000;
+    constexpr int SENTINEL = -1;
+
+    for (int rep = 0; rep < kReps; ++rep) {
+        // Small ring + tiny spin window so the ring is frequently empty exactly
+        // when the sentinel is published — the interleaving under test.
+        Channel<int> ch(/*capacity=*/4, /*spin_count=*/8);
+
+        std::thread producer([&] {
+            for (int i = 0; i < N; ++i) {
+                for (;;) {
+                    try { ch.push(i); break; }
+                    catch (const ChannelOverflowError&) { std::this_thread::yield(); }
+                }
+            }
+            ch.push_sentinel(SENTINEL);   // must-deliver, never overflows/blocks
+        });
+
+        int  expected = 0;
+        bool in_order = true;
+        bool saw_sentinel = false;
+        // Treat the sentinel as EOF: stop draining the instant it appears.
+        for (;;) {
+            int v = ch.pop();
+            if (v == SENTINEL) { saw_sentinel = true; break; }
+            if (v != expected) in_order = false;
+            ++expected;
+        }
+        producer.join();
+
+        REQUIRE(saw_sentinel);
+        REQUIRE(in_order);
+        REQUIRE(expected == N);        // all N values received before the sentinel
+        REQUIRE(ch.size() == 0);
+        REQUIRE(ch.approx_size() == 0);
+    }
+}
+
+TEST_CASE("SPSC: sentinel is strictly last, after every value (try_pop_now)",
+          "[channel][stress]") {
+    // The pool-node consume path is try_pop_now(), not pop(): it must surface
+    // the out-of-band sentinel only once the ring is freshly observed empty.
+    // The consumer spins with no sleeps, racing the producer at full tilt
+    // across the empty-ring boundary where take_sentinel() is reached.
+    constexpr int N        = 20'000;
+    constexpr int SENTINEL = -1;
+
+    for (int rep = 0; rep < kReps; ++rep) {
+        Channel<int> ch(/*capacity=*/4, /*spin_count=*/0);
+
+        std::thread producer([&] {
+            for (int i = 0; i < N; ++i) {
+                for (;;) {
+                    try { ch.push(i); break; }
+                    catch (const ChannelOverflowError&) { std::this_thread::yield(); }
+                }
+            }
+            ch.push_sentinel(SENTINEL);
+        });
+
+        int  expected = 0;
+        bool in_order = true;
+        bool saw_sentinel = false;
+        int  v;
+        for (;;) {
+            if (!ch.try_pop_now(v)) { std::this_thread::yield(); continue; }
+            if (v == SENTINEL) { saw_sentinel = true; break; }
+            if (v != expected) in_order = false;
+            ++expected;
+        }
+        producer.join();
+
+        REQUIRE(saw_sentinel);
+        REQUIRE(in_order);
+        REQUIRE(expected == N);
+        // Sentinel held no ring slot; once taken the channel is fully empty.
+        REQUIRE(ch.size() == 0);
+        REQUIRE(ch.approx_size() == 0);
+    }
+}