ci: provoke pipeline (touch Dockerfile + code)

Trivial comment changes to exercise the new CI orchestration: the
docker job should rebuild+push the builder image first, then test
and docs run against the fresh image.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.gitea/workflows/ci.yaml

@@ -0,0 +1,79 @@
+name: '🚦 CI'
+
+# Single orchestrator. This is the only workflow that triggers on push/PR.
+# It decides which reusable sub-workflows to run and in what order:
+#   changes ─┬─> docker (only if the Dockerfile/requirements changed) ─┬─> test
+#            │                                                          └─> docs
+# When the builder image is rebuilt it MUST finish (and push) before test/docs
+# run, so they validate against the fresh image.
+on:
+  push:
+    branches:
+      - master
+      - develop
+  pull_request:
+    branches:
+      - master
+      - develop
+  workflow_dispatch:
+
+jobs:
+  # Detect which parts of the repo changed in this push/PR.
+  changes:
+    runs-on: linux/amd64
+    # Runs in the builder image because the host has no Node, which the
+    # JS-based checkout/paths-filter actions require.
+    container:
+      image: gitea.tourolle.paris/dtourolle/kpnpp-builder:latest
+    outputs:
+      dockerfile: ${{ steps.filter.outputs.dockerfile }}
+      code: ${{ steps.filter.outputs.code }}
+      docs: ${{ steps.filter.outputs.docs }}
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Detect changed paths
+        id: filter
+        uses: dorny/paths-filter@v3
+        with:
+          filters: |
+            dockerfile:
+              - 'Dockerfile.builder'
+              - 'docs/requirements.txt'
+            docs:
+              - 'docs/**'
+              - 'mkdocs.yml'
+              - 'examples/**/*.cpp'
+            code:
+              - 'src/**'
+              - 'include/**'
+              - 'tests/**'
+              - 'examples/**'
+              - 'python/**'
+              - 'CMakeLists.txt'
+              - '**/*.cpp'
+              - '**/*.hpp'
+              - '**/*.h'
+
+  # Rebuild the builder image first, but only when it actually changed.
+  # On pull requests we build to validate the Dockerfile but do not push.
+  docker:
+    needs: changes
+    if: ${{ needs.changes.outputs.dockerfile == 'true' }}
+    uses: ./.gitea/workflows/docker.yaml
+    with:
+      push: ${{ github.event_name != 'pull_request' }}
+
+  # 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.
+  test:
+    needs: [changes, docker]
+    if: ${{ !failure() && !cancelled() && (needs.changes.outputs.code == 'true' || needs.changes.outputs.dockerfile == 'true') }}
+    uses: ./.gitea/workflows/test.yaml
+
+  docs:
+    needs: [changes, docker]
+    if: ${{ !failure() && !cancelled() && github.ref == 'refs/heads/master' && (needs.changes.outputs.docs == 'true' || needs.changes.outputs.dockerfile == 'true') }}
+    uses: ./.gitea/workflows/docs.yaml
+    secrets: inherit

.gitea/workflows/docker.yaml

@@ -0,0 +1,54 @@
+name: '🐳 Builder Image'
+
+# Reusable workflow: builds (and optionally pushes) the kpnpp-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.
+on:
+  workflow_call:
+    inputs:
+      push:
+        description: 'Push the built image to the registry'
+        type: boolean
+        default: true
+  workflow_dispatch:
+    inputs:
+      push:
+        description: 'Push the built image to the registry'
+        type: boolean
+        default: true
+
+jobs:
+  build:
+    runs-on: linux/amd64
+    steps:
+      # This job runs on the host (not in a container) so it can reach the
+      # host Docker daemon and reuse the cached registry credentials. The host
+      # has no Node, so the JS-based actions/checkout can't run here; do a
+      # minimal shallow fetch of this commit with plain git instead.
+      - name: Checkout repository
+        run: |
+          git init -q .
+          git remote add origin "${{ github.server_url }}/${{ github.repository }}.git"
+          git -c http.extraheader="AUTHORIZATION: basic $(printf '%s' '${{ github.actor }}:${{ github.token }}' | base64 -w0)" \
+            fetch --depth 1 origin "${{ github.sha }}"
+          git checkout -q FETCH_HEAD
+
+      # No docker login step: the host runner was authenticated to
+      # gitea.tourolle.paris with `docker login` during setup, so its cached
+      # credentials in ~/.docker/config.json cover the push below.
+      - name: Build builder image
+        # Context is the repo root because Dockerfile.builder COPYs
+        # docs/requirements.txt during the build.
+        run: |
+          docker build \
+            -f Dockerfile.builder \
+            -t gitea.tourolle.paris/dtourolle/kpnpp-builder:latest \
+            -t gitea.tourolle.paris/dtourolle/kpnpp-builder:${{ github.sha }} \
+            .
+
+      - name: Push builder image
+        if: ${{ inputs.push }}
+        run: |
+          docker push gitea.tourolle.paris/dtourolle/kpnpp-builder:latest
+          docker push gitea.tourolle.paris/dtourolle/kpnpp-builder:${{ github.sha }}

.gitea/workflows/docs.yaml

@@ -0,0 +1,34 @@
+name: '📚 Docs'
+
+# Triggering and path filtering are owned by ci.yaml (the orchestrator), which
+# calls this as a reusable workflow. workflow_dispatch is kept for manual runs.
+on:
+  workflow_call:
+  workflow_dispatch:
+
+jobs:
+  deploy:
+    runs-on: linux/amd64
+    container:
+      image: gitea.tourolle.paris/dtourolle/kpnpp-builder:latest
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0   # full history needed for mkdocs gh-deploy
+
+      - name: Configure git identity
+        run: |
+          git config user.name  "Gitea Actions"
+          git config user.email "actions@gitea.tourolle.paris"
+
+      - name: Build and deploy to gitea-pages branch
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          mkdocs gh-deploy \
+            --force \
+            --remote-branch gitea-pages \
+            --remote-name origin \
+            --message "docs: deploy from ${{ github.sha }}"

.gitea/workflows/test.yaml

@@ -1,18 +1,9 @@
 name: '🧪 Test'
 
+# Triggering and path filtering are owned by ci.yaml (the orchestrator), which
+# calls this as a reusable workflow. workflow_dispatch is kept for manual runs.
 on:
-  push:
-    branches:
-      - master
-      - develop
-    paths-ignore:
-      - '**/*.md'
-  pull_request:
-    branches:
-      - master
-      - develop
-    paths-ignore:
-      - '**/*.md'
+  workflow_call:
   workflow_dispatch:
 
 jobs:
@@ -41,7 +32,7 @@ jobs:
             -G Ninja \
             -DCMAKE_BUILD_TYPE=Debug \
             -DKPN_BUILD_TESTS=ON \
-            -DKPN_BUILD_EXAMPLES=OFF \
+            -DKPN_BUILD_EXAMPLES=ON \
             -DKPN_BUILD_PYTHON=ON \
             -DFETCHCONTENT_BASE_DIR=$HOME/.cmake/fetchcontent
 
@@ -49,18 +40,26 @@ jobs:
         working-directory: test-${{ github.run_id }}
         run: cmake --build build --parallel
 
-      - name: Run tests
+      - name: Run unit tests
         working-directory: test-${{ github.run_id }}
         run: |
           cd build
-          ctest --output-on-failure --output-junit test-results.xml
+          ctest --output-on-failure --output-junit test-results.xml --label-exclude examples
+
+      - name: Run example smoke tests
+        working-directory: test-${{ github.run_id }}
+        run: |
+          cd build
+          ctest --output-on-failure --output-junit example-results.xml -L examples
 
       - name: Upload test results
         if: always()
         uses: actions/upload-artifact@v3
         with:
           name: test-results
-          path: test-${{ github.run_id }}/build/test-results.xml
+          path: |
+            test-${{ github.run_id }}/build/test-results.xml
+            test-${{ github.run_id }}/build/example-results.xml
           retention-days: 7
 
       - name: Cleanup

.gitignore

@@ -1,6 +1,7 @@
 # Build output
 build/
-
+build_debug/
+site/
 # Python
 __pycache__/
 *.py[cod]

CMakeLists.txt

@@ -48,6 +48,12 @@ if(KPN_BUILD_TESTS)
     add_subdirectory(tests)
 endif()
 
+# ── Benchmarks ────────────────────────────────────────────────────────────────
+option(KPN_BUILD_BENCHMARKS "Build benchmarks" OFF)
+if(KPN_BUILD_BENCHMARKS)
+    add_subdirectory(benchmarks)
+endif()
+
 # ── Python bindings ───────────────────────────────────────────────────────────
 if(KPN_BUILD_PYTHON)
     find_package(Python 3.8 COMPONENTS Interpreter Development.Module REQUIRED)
@@ -69,3 +75,14 @@ endif()
 if(KPN_BUILD_EXAMPLES)
     add_subdirectory(examples)
 endif()
+
+# ── Docs (README generation) ──────────────────────────────────────────────────
+find_package(Python3 QUIET COMPONENTS Interpreter)
+if(Python3_FOUND)
+    add_custom_target(docs
+        COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/scripts/render_readme.py
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        COMMENT "Rendering README.md from README.md.in"
+        VERBATIM
+    )
+endif()

Dockerfile.builder

@@ -1,4 +1,4 @@
-# KPN++ Builder Image
+# KPN++ Builder Image  (CI: pipeline trigger)
 # 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
@@ -16,4 +16,12 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     nodejs \
     && rm -rf /var/lib/apt/lists/*
 
+# Pre-install MkDocs dependencies so the docs workflow does not need to pip
+# install at runtime. --break-system-packages is required because the Debian
+# base marks the environment as externally managed (PEP 668); this is safe in
+# a dedicated container image.
+COPY docs/requirements.txt /tmp/docs-requirements.txt
+RUN pip install --no-cache-dir --break-system-packages -r /tmp/docs-requirements.txt \
+    && rm /tmp/docs-requirements.txt
+
 WORKDIR /src

README.md

@@ -50,31 +50,55 @@ A node wraps any callable. Its input types are taken from the function's paramet
 ```cpp
 #include <kpn/kpn.hpp>
 using namespace kpn;
+```
 
-// Single input, single output
-int double_it(int x) { return x * 2; }
+Source, transform, and sink — from [`examples/01_hello_pipeline/main.cpp`](examples/01_hello_pipeline/main.cpp):
 
-// Multi-output — must return std::tuple
-std::tuple<cv::Mat, cv::Mat> split(cv::Mat frame) { ... }
+```cpp
+static int  produce()        { return 42; }
+static int  double_it(int x) { return x * 2; }
+static void print_it(int x)  { std::cout << "result: " << x << '\n'; }
+```
 
-// Sink — void return, no output ports
-void display(cv::Mat frame) { cv::imshow("out", frame); }
+Multi-output node returning a tuple — from [`examples/03_multi_output/main.cpp`](examples/03_multi_output/main.cpp):
+
+```cpp
+// Multi-output: returns (key, value) as a tuple — KPN++ routes each element
+// to its own output port automatically.
+static std::tuple<std::string, std::string> parse(std::string kv) {
+    auto sep = kv.find('=');
+    if (sep == std::string::npos) return {kv, ""};
+    return {kv.substr(0, sep), kv.substr(sep + 1)};
+}
 ```
 
 ### Creating Nodes
 
+**Index-only ports** (from [`examples/01_hello_pipeline/main.cpp`](examples/01_hello_pipeline/main.cpp)):
+
 ```cpp
-// No port names (index-only access)
-auto node = make_node<double_it>(/*fifo_capacity=*/5);
+auto src  = make_node<produce>(5);
+auto dbl  = make_node<double_it>(5);
+auto sink = make_node<print_it>(5);
+```
 
-// Named input ports only
-auto node = make_node<double_it>(in<"value">{}, 5);
+**Named ports** (from [`examples/02_named_ports/main.cpp`](examples/02_named_ports/main.cpp)):
 
-// Named input and output ports
-auto node = make_node<double_it>(in<"value">{}, out<"result">{}, 5);
+```cpp
+// tokenise: no inputs, one named output "words"
+auto tok = make_node<tokenise>(out<"words">{}, 4);
 
-// Named output ports only (e.g. a source with no inputs)
-auto node = make_node<capture>(out<"colour","grey">{}, 5);
+// count_words: named input "words", named outputs "count" and "words"
+auto cnt = make_node<count_words>(in<"words">{}, out<"count", "words">{}, 4);
+
+// report: two named inputs
+auto snk = make_node<report>(in<"count", "words">{}, 4);
+```
+
+**Multi-named output source** (from [`examples/09_opencv_cellshade/main.cpp`](examples/09_opencv_cellshade/main.cpp)):
+
+```cpp
+auto src = make_node<capture>(out<"colour","grey">{}, 8);
 ```
 
 Port names are NTTP `fixed_string` values — resolved entirely at compile time, zero runtime cost.
@@ -83,22 +107,25 @@ Port names are NTTP `fixed_string` values — resolved entirely at compile time,
 
 `Network` is **non-owning** — declare nodes first, then register them. Nodes must outlive the network.
 
-```cpp
-auto src  = make_node<produce>(in<"x">{}, out<"value">{}, 5);
-auto proc = make_node<double_it>(in<"value">{}, out<"result">{}, 5);
+From [`examples/02_named_ports/main.cpp`](examples/02_named_ports/main.cpp):
 
+```cpp
 Network net;
-net.add("src",  src)
-   .add("proc", proc)
-   .connect("src",  src.output<0>(),           "proc", proc.input<0>())   // by index
-   .connect("src",  src.template output<"value">(), "proc", proc.template input<"value">())  // by name
-   .build();   // runs cycle detection — throws NetworkCycleError on cycles
+net.add("tok", tok)
+   .add("cnt", cnt)
+   .add("snk", snk)
+   .connect("tok", tok.template output<"words">(), "cnt", cnt.template input<"words">())
+   .connect("cnt", cnt.template output<"count">(), "snk", snk.template input<"count">())
+   .connect("cnt", cnt.template output<"words">(), "snk", snk.template input<"words">())
+   .build();
 
 net.start();
-// ... do work ...
+std::this_thread::sleep_for(std::chrono::milliseconds(500));
 net.stop();
 ```
 
+`.build()` runs cycle detection — throws `NetworkCycleError` on cycles.
+
 > **Named port syntax in template context:** when the node variable is `auto`-deduced, use `.template output<"name">()` and `.template input<"name">()` to help the parser.
 
 ### Channel Semantics
@@ -113,14 +140,36 @@ net.stop();
 
 Large types (`sizeof > 8` or non-trivially-copyable) are stored as `std::shared_ptr<const T>` inside the channel — no copies, shared immutable ownership. Small trivially-copyable types are stored by value.
 
-Override the policy for a specific type:
+Override the policy for a specific type (from [`examples/04_storage_policy/main.cpp`](examples/04_storage_policy/main.cpp)):
 
 ```cpp
-template<> struct kpn::channel_storage_policy<MyType> {
+// Override: store Tag by value despite being a struct
+// (it's trivially copyable and small — this just makes the policy explicit)
+template<>
+struct kpn::channel_storage_policy<Tag> {
     static constexpr bool by_value = true;
 };
 ```
 
+### Diagnostics & Error Handling
+
+Custom diagnostics handler — fires on the watchdog interval (from [`examples/05_error_handling/main.cpp`](examples/05_error_handling/main.cpp)):
+
+```cpp
+// Custom diagnostics handler — fires on the watchdog interval.
+// Print a concise one-liner rather than the full table.
+net.set_diagnostics_handler([](const std::vector<NodeSnapshot>& nodes,
+                               const std::vector<ChannelSnapshot>& channels) {
+    std::cout << "[diag] ";
+    for (auto& n : nodes)
+        std::cout << n.name << "=" << n.throughput_fps << "fps  ";
+    for (auto& c : channels)
+        std::cout << "channel fill=" << static_cast<int>(c.fill_pct()) << "% "
+                  << "overflows=" << c.overflows;
+    std::cout << '\n';
+});
+```
+
 ### Shutdown
 
 `node.stop()` / `net.stop()`:
@@ -171,20 +220,155 @@ top.start();
 
 ## Display / GUI Nodes
 
-**Do not wrap `imshow`/`waitKey` as a KPN node.** Qt and Wayland require these to run on the main thread (the thread that owns the event loop). Instead, wire the final output channel to a `Channel<cv::Mat>` and pop it on the main thread:
+**Do not wrap `imshow`/`waitKey` as a KPN node.** Qt and Wayland require these to run on the main thread (the thread that owns the event loop). Instead, derive from `MainThreadNode<>` — it owns the input channels, implements `INode`, and exposes a `step()` method to call on the main thread.
+
+`DisplayNode` from [`examples/09_opencv_cellshade/main.cpp`](examples/09_opencv_cellshade/main.cpp):
 
 ```cpp
-Channel<cv::Mat> result_ch(8);
-comp.set_output_channel<0>(&result_ch);
-// ... build and start network ...
+class DisplayNode : public kpn::MainThreadNode<DisplayNode,
+                                               kpn::in<"composite", "edges">,
+                                               cv::Mat, cv::Mat> {
+public:
+    DisplayNode() : MainThreadNode(8) {
+        cv::namedWindow("Cell Shade", cv::WINDOW_NORMAL);
+        cv::namedWindow("Edge Mask",  cv::WINDOW_NORMAL);
+        cv::resizeWindow("Cell Shade", 1280, 720);
+        cv::resizeWindow("Edge Mask",   640, 360);
+    }
+
+    ~DisplayNode() { cv::destroyAllWindows(); }
+
+    bool operator()(cv::Mat composite, cv::Mat edges) {
+        cv::imshow("Cell Shade", composite);
+        cv::Mat edges_bgr;
+        cv::cvtColor(edges, edges_bgr, cv::COLOR_GRAY2BGR);
+        cv::imshow("Edge Mask", edges_bgr);
+        int key = cv::waitKey(1);
+        if (key == 'q' || key == 27) return false;
+        return window_open("Cell Shade") && window_open("Edge Mask");
+    }
+
+private:
+    static bool window_open(const char* name) {
+        try { return cv::getWindowProperty(name, cv::WND_PROP_VISIBLE) >= 1; }
+        catch (const cv::Exception&) { return false; }
+    }
+};
+```
+
+Wire it into the network and drive it from the main thread:
+
+```cpp
+net.start();
+
+// Main thread drives display — imshow/waitKey stay on the GUI thread.
+// step() returns false when operator() returns false (q pressed / window closed).
+while (disp.step())
+    cv::waitKey(8); // yield event loop when no frame ready
+
+net.stop();
+```
+
+---
+
+## OpenCV Cell-Shading Example
+
+Real-time cell-shading pipeline from [`examples/09_opencv_cellshade/main.cpp`](examples/09_opencv_cellshade/main.cpp).
+
+**Source node** — returns two frames (colour + grey) as a tuple, routing them to separate downstream branches:
+
+```cpp
+static std::tuple<cv::Mat, cv::Mat> capture() {
+    constexpr int W = 640, H = 480;
+    static cv::VideoCapture cap;
+    static bool opened = false;
+    if (!opened) {
+        opened = true;
+        cap.open(0, cv::CAP_V4L2);
+        if (cap.isOpened()) {
+            cap.set(cv::CAP_PROP_FRAME_WIDTH,  W);
+            cap.set(cv::CAP_PROP_FRAME_HEIGHT, H);
+        } else {
+            std::cerr << "[capture] no webcam — using synthetic animated pattern\n";
+        }
+    }
 
-// Main thread display loop
-while (true) {
     cv::Mat frame;
-    if (!result_ch.try_pop(frame, std::chrono::milliseconds(100))) continue;
-    cv::imshow("output", frame);
-    if (cv::waitKey(1) == 'q') break;
+    if (cap.isOpened()) {
+        auto t0 = std::chrono::steady_clock::now();
+        cap >> frame;
+        auto elapsed = std::chrono::steady_clock::now() - t0;
+        if (elapsed < std::chrono::milliseconds(20))
+            std::this_thread::sleep_for(std::chrono::milliseconds(33) - elapsed);
+        if (frame.empty()) frame = cv::Mat::zeros(H, W, CV_8UC3);
+    } else {
+        static int tick = 0;
+        static cv::Mat grad = make_gradient(W, H);
+        ++tick;
+        frame = grad.clone();
+        int r = 150 + (tick % 80) * 4;
+        cv::circle(frame, {W/2,   H/2},   r,     {255, 200,   0}, -1);
+        cv::circle(frame, {W/2,   H/2},   r / 2, {  0, 128, 255}, -1);
+        cv::circle(frame, {W*2/5, H*2/5}, r / 3, {200,   0, 200}, -1);
+        std::this_thread::sleep_for(std::chrono::milliseconds(33));
+    }
+    return {frame.clone(), frame.clone()};
 }
+```
+
+**Full network wiring:**
+
+```cpp
+auto src       = make_node<capture>  (out<"colour","grey">{},            8);
+auto gray_node = make_node<to_gray>  (in<"bgr">{},   out<"gray">{},      8);
+auto edge_node = make_node<edges_fn> (in<"gray">{},  out<"edges">{},     8);
+auto quant     = make_node<quantise> (in<"bgr">{},   out<"quantised">{}, 8);
+auto comp      = make_node<composite>(in<"edges","colour">{}, out<"result","edges">{}, 8);
+
+// DisplayNode: two windows opened in constructor, step() drives main thread.
+DisplayNode disp;
+
+Network net;
+net.add("src",     src)
+   .add("gray",    gray_node)
+   .add("edges",   edge_node)
+   .add("quant",   quant)
+   .add("comp",    comp)
+   .add("display", disp)
+   .connect("src",   src.template output<"colour">(),      "quant",   quant.template input<"bgr">())
+   .connect("quant", quant.template output<"quantised">(), "comp",    comp.template input<"colour">())
+   .connect("src",   src.template output<"grey">(),        "gray",    gray_node.template input<"bgr">())
+   .connect("gray",  gray_node.template output<"gray">(),  "edges",   edge_node.template input<"gray">())
+   .connect("edges", edge_node.template output<"edges">(), "comp",    comp.template input<"edges">())
+   .connect("comp",  comp.template output<"result">(),     "display", disp.template input<"composite">())
+   .connect("comp",  comp.template output<"edges">(),      "display", disp.template input<"edges">())
+   .build();
+```
+
+---
+
+## Fan-Out (Multi-Output)
+
+From [`examples/03_multi_output/main.cpp`](examples/03_multi_output/main.cpp) — one node fans out to two independent sinks via a tuple return:
+
+```cpp
+auto gen  = make_node<generate>(out<"kv">{},          4);
+auto par  = make_node<parse>   (in<"kv">{},  out<"key", "value">{}, 4);
+auto keys = make_node<print_key>  (in<"key">{},   4);
+auto vals = make_node<print_value>(in<"value">{}, 4);
+
+Network net;
+net.add("gen",  gen)
+   .add("par",  par)
+   .add("keys", keys)
+   .add("vals", vals)
+   .connect("gen",  gen.template output<"kv">(),    "par",  par.template input<"kv">())
+   .connect("par",  par.template output<"key">(),   "keys", keys.template input<"key">())
+   .connect("par",  par.template output<"value">(), "vals", vals.template input<"value">())
+   .build();
+
+net.start();
+std::this_thread::sleep_for(std::chrono::milliseconds(600));
 net.stop();
 ```
 
@@ -228,6 +412,143 @@ Run the cell-shading example:
 
 ---
 
+## Performance
+
+Measured on Linux (x86-64, `-O3 -march=native`) with `benchmarks/bench_pipeline`.
+Each topology pushes N items through the graph; `overhead_us/item` strips out the
+per-node compute time to isolate framework cost.
+
+Overhead formula: `(elapsed − (N + depth − 1) × work_us) / N` removes the expected
+pipeline-fill cost so the number reflects pure framework latency.
+
+### Baseline overhead (private pools, 100 µs/node)
+
+| Topology | items/sec | overhead µs/item |
+|---|---|---|
+| chain depth-1 | 9 797 | ~2 |
+| chain depth-4 | 9 448 | ~4 |
+| chain depth-8 | 9 078 | ~7 |
+| chain depth-16 | 7 004 | ~13 ← oversubscription |
+| chain depth-32 | 4 179 | ~77 ← oversubscription |
+| wide fanout-1 | 9 751 | ~3 |
+| wide fanout-4 | 9 668 | ~3 |
+| diamond (2×2) | 9 607 | ~4 |
+
+Chain overhead is flat at **~2–7 µs/hop** for depths within the machine's core count,
+then rises once threads compete for CPU. Wide and diamond topologies add no measurable
+overhead as fanout increases — all branches run in parallel.
+
+### Scheduling modes
+
+`Node<>` gives each node a private `ThreadPool(1)`. `PoolNode<>` lets multiple
+nodes share one pool. The right choice depends on the graph shape:
+
+| Scenario | Recommended |
+|---|---|
+| Work per node < 100 µs, deep chain | Private pools — lower per-hop latency |
+| Work per node ≥ 100 µs, wide/diamond | Shared pool, `threads = hardware_concurrency` |
+| Any graph, bounded thread count required | Shared pool, `threads ≥ max parallel nodes` |
+
+A shared single-thread pool (`threads=1`) fully serialises the graph — throughput
+divides by depth for chains and by width for fanout topologies. A shared pool with
+`threads ≥ max_concurrent_nodes` matches private-pool throughput while keeping the
+OS thread count bounded.
+
+### vs. TBB flow graph
+
+Benchmarked against `tbb::flow::function_node<int,int>` (serial concurrency) with
+`tbb::flow::broadcast_node<int>` for fanout. Run with `cmake -DKPN_BUILD_BENCHMARKS=ON`
+— TBB benchmarks are included automatically when `find_package(TBB)` succeeds.
+
+**Channel implementation:** lock-free SPSC ring buffer with `std::atomic::wait/notify_one`
+(C++20 portable futex) plus a configurable spin-before-sleep window (default ~4 µs).
+Large types are stored as `shared_ptr<const T>` — fanout copies reference counts,
+not data.
+
+Overhead µs/item at **work_us = 10** (framework overhead dominates):
+
+| Topology | KPN private | TBB |
+|---|---|---|
+| chain depth-1 | 1.7 | **1.4** |
+| chain depth-4 | 2.5 | **2.2** |
+| chain depth-8 | **3.0** | 3.6 |
+| chain depth-16 | **9.3** | 13.0 |
+| chain depth-32 | 23.2 | **14.2** |
+| wide fanout-4 | 2.5 | **1.4** |
+| diamond (2×2) | 3.4 | **1.9** |
+
+Overhead µs/item at **work_us = 100** (moderate compute, KPN wins):
+
+| Topology | KPN private | TBB |
+|---|---|---|
+| chain depth-1 | **2.1** | 3.5 |
+| chain depth-4 | **4.3** | 5.2 |
+| chain depth-8 | **6.7** | 8.5 |
+| chain depth-16 | **12.8** | 17.4 |
+| chain depth-32 | **77** | 81 |
+| wide fanout-4 | 3.4 | **1.9** |
+| diamond (2×2) | **4.1** | 6.1 |
+
+KPN private pools beat TBB for every chain and diamond topology at 100 µs/node, and
+match TBB within ~20% at 10 µs/node for shallow chains. TBB retains an edge on wide
+fanout (serial dispatch loop vs. work-stealing pool) and at extreme oversubscription
+depths (chain-32 at 10 µs). The remaining gap at light work is the cost of
+`atomic::wait` vs. TBB's continuously-spinning worker threads.
+
+### vs. TBB — API
+
+The function signature is the node. KPN infers input and output types automatically;
+there is no graph object to manage.
+
+**Single-output node:**
+```cpp
+// KPN — 1 line
+int scale(int x) { return x * 2; }
+
+// TBB — must state types, concurrency policy, and carry a graph reference
+tbb::flow::function_node<int,int> n(g, tbb::flow::serial, [](int x){ return x*2; });
+```
+
+**Multi-output node:**
+```cpp
+// KPN — return a tuple
+std::tuple<cv::Mat,cv::Mat> split(cv::Mat f) { return {f, f}; }
+
+// TBB — multifunction_node + explicit try_put per port
+tbb::flow::multifunction_node<cv::Mat, std::tuple<cv::Mat,cv::Mat>> n(
+    g, tbb::flow::serial,
+    [](cv::Mat f, auto& ports) {
+        std::get<0>(ports).try_put(f);
+        std::get<1>(ports).try_put(f);
+    });
+```
+
+**Named ports** — compile-time checked, zero runtime cost, not available in TBB:
+```cpp
+auto node = make_node<split>(in<"frame">{}, out<"colour","grey">{}, 5);
+net.connect("cam", cam.output<"frame">(), "split", node.input<"frame">());
+//                                                         ^^^^^^^ typo → compile error
+```
+
+| | KPN | TBB |
+|---|---|---|
+| Node definition | plain function | `function_node<In,Out>` + explicit types |
+| Multi-output | `return std::tuple<A,B>` | `multifunction_node` + `try_put` × N |
+| Named ports | `in<"name">` / `out<"name">` compile-time | none |
+| Graph lifetime | none | `graph g` must outlive all nodes |
+| Shutdown | `net.stop()` | `g.wait_for_all()` + manual |
+| Python bindings | designed-in | none |
+
+Build the benchmarks with:
+
+```bash
+cmake -B build -DKPN_BUILD_BENCHMARKS=ON
+cmake --build build --target bench_pipeline
+./build/benchmarks/bench_pipeline | tee results.csv
+```
+
+---
+
 ## Project Structure
 
 ```
@@ -254,4 +575,6 @@ python/
   kpn_python.cpp     — nanobind module entry point
 examples/
   01_hello_pipeline/ … 09_opencv_cellshade/
+scripts/
+  render_readme.py   — regenerates README.md from README.md.in
 ```

README.md.in

@@ -0,0 +1,402 @@
+# KPN++
+
+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.
+
+---
+
+## Requirements
+
+| Dependency | Version | Notes |
+|---|---|---|
+| CMake | ≥ 3.21 | |
+| C++ compiler | GCC ≥ 11, Clang ≥ 13, MSVC 19.29 | C++20 required |
+| Threads | system | `find_package(Threads)` |
+| nanobind | ≥ 2.1 | auto-fetched if not installed; Python ≥ 3.8 |
+| Catch2 | v3 | auto-fetched for tests |
+| Google Test | v1.14 | auto-fetched for tests |
+| OpenCV | ≥ 4 | optional; only for example 09 |
+
+---
+
+## Build
+
+```bash
+cmake -B build -DKPN_BUILD_PYTHON=OFF   # core + tests + C++ examples
+cmake --build build --parallel
+ctest --test-dir build
+```
+
+Enable Python bindings (requires nanobind and Python dev headers):
+
+```bash
+cmake -B build -DKPN_BUILD_PYTHON=ON
+cmake --build build --parallel
+```
+
+Disable examples:
+
+```bash
+cmake -B build -DKPN_BUILD_EXAMPLES=OFF
+```
+
+---
+
+## Core Concepts
+
+### Nodes
+
+A node wraps any callable. Its input types are taken from the function's parameter list; its output types from the return type. Multi-output nodes return `std::tuple<...>`.
+
+```cpp
+#include <kpn/kpn.hpp>
+using namespace kpn;
+```
+
+Source, transform, and sink — from [`examples/01_hello_pipeline/main.cpp`](examples/01_hello_pipeline/main.cpp):
+
+<!-- @snippet examples/01_hello_pipeline/main.cpp basic_node_fns -->
+
+Multi-output node returning a tuple — from [`examples/03_multi_output/main.cpp`](examples/03_multi_output/main.cpp):
+
+<!-- @snippet examples/03_multi_output/main.cpp multi_output_fn -->
+
+### Creating Nodes
+
+**Index-only ports** (from [`examples/01_hello_pipeline/main.cpp`](examples/01_hello_pipeline/main.cpp)):
+
+<!-- @snippet examples/01_hello_pipeline/main.cpp index_only_nodes -->
+
+**Named ports** (from [`examples/02_named_ports/main.cpp`](examples/02_named_ports/main.cpp)):
+
+<!-- @snippet examples/02_named_ports/main.cpp named_port_creation -->
+
+**Multi-named output source** (from [`examples/09_opencv_cellshade/main.cpp`](examples/09_opencv_cellshade/main.cpp)):
+
+```cpp
+auto src = make_node<capture>(out<"colour","grey">{}, 8);
+```
+
+Port names are NTTP `fixed_string` values — resolved entirely at compile time, zero runtime cost.
+
+### Building a Network
+
+`Network` is **non-owning** — declare nodes first, then register them. Nodes must outlive the network.
+
+From [`examples/02_named_ports/main.cpp`](examples/02_named_ports/main.cpp):
+
+<!-- @snippet examples/02_named_ports/main.cpp named_port_network -->
+
+`.build()` runs cycle detection — throws `NetworkCycleError` on cycles.
+
+> **Named port syntax in template context:** when the node variable is `auto`-deduced, use `.template output<"name">()` and `.template input<"name">()` to help the parser.
+
+### Channel Semantics
+
+- **Bounded FIFO**: default capacity 5, configurable per-node at construction.
+- **Blocking `pop()`**: consumer blocks until data is available (KPN semantics).
+- **Throwing `push()`**: throws `ChannelOverflowError` if the channel is full and accepting.
+- **Silent drop on disabled channel**: after `node.stop()`, its input channels are disabled — producers that push into them have the value silently dropped. No exception, no blocking.
+- **Source throttling**: source nodes (no inputs) must sleep or yield to avoid overflowing downstream FIFOs. See example 09.
+
+### Storage Policy
+
+Large types (`sizeof > 8` or non-trivially-copyable) are stored as `std::shared_ptr<const T>` inside the channel — no copies, shared immutable ownership. Small trivially-copyable types are stored by value.
+
+Override the policy for a specific type (from [`examples/04_storage_policy/main.cpp`](examples/04_storage_policy/main.cpp)):
+
+<!-- @snippet examples/04_storage_policy/main.cpp storage_policy_spec -->
+
+### Diagnostics & Error Handling
+
+Custom diagnostics handler — fires on the watchdog interval (from [`examples/05_error_handling/main.cpp`](examples/05_error_handling/main.cpp)):
+
+<!-- @snippet examples/05_error_handling/main.cpp diagnostics_handler -->
+
+### Shutdown
+
+`node.stop()` / `net.stop()`:
+1. Sets `accepting_ = false` on all input channels (drops in-flight pushes silently).
+2. Clears any queued items from those channels.
+3. Unblocks any thread blocked on `pop()` (throws `ChannelClosedError` inside `run_loop`, which exits cleanly).
+4. Joins the node thread.
+
+---
+
+## Named Ports — Design Notes
+
+Port names use C++20 NTTP `fixed_string`. The deduction guide is required:
+
+```cpp
+template<std::size_t N>
+fixed_string(const char (&)[N]) -> fixed_string<N>;
+```
+
+`fixed_string<4>` and `fixed_string<7>` are distinct types — `input<"img">()` and `input<"sigma">()` resolve to different template instantiations at compile time. Wrong names produce a `static_assert` at the call site with a readable message.
+
+---
+
+## Sub-Networks
+
+`Network` implements `INode`, so it can be nested inside a larger `Network`:
+
+```cpp
+// Inner sub-network
+Network pipe;
+pipe.add("pre", pre_node)
+    .add("enh", enh_node)
+    .connect("pre", pre_node.output<0>(), "enh", enh_node.input<0>())
+    .expose_input("img",    pre_node.input<0>())
+    .expose_output("result", enh_node.output<0>())
+    .build();
+
+// Outer network
+Network top;
+top.add("pipe", pipe)
+   .add("sink", sink_node)
+   .connect("pipe", pipe.output<"result">(), "sink", sink_node.input<0>())
+   .build();
+top.start();
+```
+
+---
+
+## Display / GUI Nodes
+
+**Do not wrap `imshow`/`waitKey` as a KPN node.** Qt and Wayland require these to run on the main thread (the thread that owns the event loop). Instead, derive from `MainThreadNode<>` — it owns the input channels, implements `INode`, and exposes a `step()` method to call on the main thread.
+
+`DisplayNode` from [`examples/09_opencv_cellshade/main.cpp`](examples/09_opencv_cellshade/main.cpp):
+
+<!-- @snippet examples/09_opencv_cellshade/main.cpp display_node -->
+
+Wire it into the network and drive it from the main thread:
+
+<!-- @snippet examples/09_opencv_cellshade/main.cpp main_thread_step -->
+
+---
+
+## OpenCV Cell-Shading Example
+
+Real-time cell-shading pipeline from [`examples/09_opencv_cellshade/main.cpp`](examples/09_opencv_cellshade/main.cpp).
+
+**Source node** — returns two frames (colour + grey) as a tuple, routing them to separate downstream branches:
+
+<!-- @snippet examples/09_opencv_cellshade/main.cpp capture_fn -->
+
+**Full network wiring:**
+
+<!-- @snippet examples/09_opencv_cellshade/main.cpp opencv_network -->
+
+---
+
+## Fan-Out (Multi-Output)
+
+From [`examples/03_multi_output/main.cpp`](examples/03_multi_output/main.cpp) — one node fans out to two independent sinks via a tuple return:
+
+<!-- @snippet examples/03_multi_output/main.cpp fanout_network -->
+
+---
+
+## Python Bindings
+
+> Python bindings are scaffolded but not yet fully implemented. See `python/kpn_python.cpp` and `include/kpn/python/bindings.hpp`.
+
+A `PyNetwork` is constructed from a closed list of C++ node types. The variant of all port types is derived at compile time — no runtime type registration needed.
+
+**GIL rules (non-negotiable):**
+- Acquire the GIL only for the duration of a Python callable invocation.
+- Release the GIL before any blocking channel operation (`pop()`, `push()`, `net.read()`, `net.write()`).
+
+Violating the second rule deadlocks.
+
+---
+
+## Examples
+
+| Example | What it shows |
+|---|---|
+| `01_hello_pipeline` | Linear pipeline, index-based port wiring |
+| `02_named_ports` | `in<>`/`out<>` name tags, named port access |
+| `03_multi_output` | Tuple-returning node, per-element sub-port routing |
+| `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)* |
+| `08_python_subport` | `net.read`, `net.write`, sub-port tap *(pending)* |
+| `09_opencv_cellshade` | Real-time cell-shading on webcam/pattern; requires OpenCV ≥ 4 |
+
+Run the cell-shading example:
+
+```bash
+./build/examples/09_opencv_cellshade
+# Press 'q' or close the window to stop.
+# Falls back to an animated synthetic pattern if no webcam is found.
+```
+
+---
+
+## Performance
+
+Measured on Linux (x86-64, `-O3 -march=native`) with `benchmarks/bench_pipeline`.
+Each topology pushes N items through the graph; `overhead_us/item` strips out the
+per-node compute time to isolate framework cost.
+
+Overhead formula: `(elapsed − (N + depth − 1) × work_us) / N` removes the expected
+pipeline-fill cost so the number reflects pure framework latency.
+
+### Baseline overhead (private pools, 100 µs/node)
+
+| Topology | items/sec | overhead µs/item |
+|---|---|---|
+| chain depth-1 | 9 797 | ~2 |
+| chain depth-4 | 9 448 | ~4 |
+| chain depth-8 | 9 078 | ~7 |
+| chain depth-16 | 7 004 | ~13 ← oversubscription |
+| chain depth-32 | 4 179 | ~77 ← oversubscription |
+| wide fanout-1 | 9 751 | ~3 |
+| wide fanout-4 | 9 668 | ~3 |
+| diamond (2×2) | 9 607 | ~4 |
+
+Chain overhead is flat at **~2–7 µs/hop** for depths within the machine's core count,
+then rises once threads compete for CPU. Wide and diamond topologies add no measurable
+overhead as fanout increases — all branches run in parallel.
+
+### Scheduling modes
+
+`Node<>` gives each node a private `ThreadPool(1)`. `PoolNode<>` lets multiple
+nodes share one pool. The right choice depends on the graph shape:
+
+| Scenario | Recommended |
+|---|---|
+| Work per node < 100 µs, deep chain | Private pools — lower per-hop latency |
+| Work per node ≥ 100 µs, wide/diamond | Shared pool, `threads = hardware_concurrency` |
+| Any graph, bounded thread count required | Shared pool, `threads ≥ max parallel nodes` |
+
+A shared single-thread pool (`threads=1`) fully serialises the graph — throughput
+divides by depth for chains and by width for fanout topologies. A shared pool with
+`threads ≥ max_concurrent_nodes` matches private-pool throughput while keeping the
+OS thread count bounded.
+
+### vs. TBB flow graph
+
+Benchmarked against `tbb::flow::function_node<int,int>` (serial concurrency) with
+`tbb::flow::broadcast_node<int>` for fanout. Run with `cmake -DKPN_BUILD_BENCHMARKS=ON`
+— TBB benchmarks are included automatically when `find_package(TBB)` succeeds.
+
+**Channel implementation:** lock-free SPSC ring buffer with `std::atomic::wait/notify_one`
+(C++20 portable futex) plus a configurable spin-before-sleep window (default ~4 µs).
+Large types are stored as `shared_ptr<const T>` — fanout copies reference counts,
+not data.
+
+Overhead µs/item at **work_us = 10** (framework overhead dominates):
+
+| Topology | KPN private | TBB |
+|---|---|---|
+| chain depth-1 | 1.7 | **1.4** |
+| chain depth-4 | 2.5 | **2.2** |
+| chain depth-8 | **3.0** | 3.6 |
+| chain depth-16 | **9.3** | 13.0 |
+| chain depth-32 | 23.2 | **14.2** |
+| wide fanout-4 | 2.5 | **1.4** |
+| diamond (2×2) | 3.4 | **1.9** |
+
+Overhead µs/item at **work_us = 100** (moderate compute, KPN wins):
+
+| Topology | KPN private | TBB |
+|---|---|---|
+| chain depth-1 | **2.1** | 3.5 |
+| chain depth-4 | **4.3** | 5.2 |
+| chain depth-8 | **6.7** | 8.5 |
+| chain depth-16 | **12.8** | 17.4 |
+| chain depth-32 | **77** | 81 |
+| wide fanout-4 | 3.4 | **1.9** |
+| diamond (2×2) | **4.1** | 6.1 |
+
+KPN private pools beat TBB for every chain and diamond topology at 100 µs/node, and
+match TBB within ~20% at 10 µs/node for shallow chains. TBB retains an edge on wide
+fanout (serial dispatch loop vs. work-stealing pool) and at extreme oversubscription
+depths (chain-32 at 10 µs). The remaining gap at light work is the cost of
+`atomic::wait` vs. TBB's continuously-spinning worker threads.
+
+### vs. TBB — API
+
+The function signature is the node. KPN infers input and output types automatically;
+there is no graph object to manage.
+
+**Single-output node:**
+```cpp
+// KPN — 1 line
+int scale(int x) { return x * 2; }
+
+// TBB — must state types, concurrency policy, and carry a graph reference
+tbb::flow::function_node<int,int> n(g, tbb::flow::serial, [](int x){ return x*2; });
+```
+
+**Multi-output node:**
+```cpp
+// KPN — return a tuple
+std::tuple<cv::Mat,cv::Mat> split(cv::Mat f) { return {f, f}; }
+
+// TBB — multifunction_node + explicit try_put per port
+tbb::flow::multifunction_node<cv::Mat, std::tuple<cv::Mat,cv::Mat>> n(
+    g, tbb::flow::serial,
+    [](cv::Mat f, auto& ports) {
+        std::get<0>(ports).try_put(f);
+        std::get<1>(ports).try_put(f);
+    });
+```
+
+**Named ports** — compile-time checked, zero runtime cost, not available in TBB:
+```cpp
+auto node = make_node<split>(in<"frame">{}, out<"colour","grey">{}, 5);
+net.connect("cam", cam.output<"frame">(), "split", node.input<"frame">());
+//                                                         ^^^^^^^ typo → compile error
+```
+
+| | KPN | TBB |
+|---|---|---|
+| Node definition | plain function | `function_node<In,Out>` + explicit types |
+| Multi-output | `return std::tuple<A,B>` | `multifunction_node` + `try_put` × N |
+| Named ports | `in<"name">` / `out<"name">` compile-time | none |
+| Graph lifetime | none | `graph g` must outlive all nodes |
+| Shutdown | `net.stop()` | `g.wait_for_all()` + manual |
+| Python bindings | designed-in | none |
+
+Build the benchmarks with:
+
+```bash
+cmake -B build -DKPN_BUILD_BENCHMARKS=ON
+cmake --build build --target bench_pipeline
+./build/benchmarks/bench_pipeline | tee results.csv
+```
+
+---
+
+## Project Structure
+
+```
+include/kpn/
+  fixed_string.hpp   — NTTP string, in<>/out<> tags, index_of
+  traits.hpp         — function_traits, normalised_return_t, output_count_v
+  channel.hpp        — Channel<T>, channel_storage_policy, exceptions
+  port.hpp           — InputPort<N,I>, OutputPort<N,I>
+  node.hpp           — Node<Func,in<...>,out<...>>, make_node, INode
+  network.hpp        — Network (builder, cycle detection, watchdog)
+  variant_node.hpp   — VariantNode, PythonConverter<T>, unique_types (Python layer)
+  python/
+    bindings.hpp     — nanobind helpers, GIL rule documentation
+  kpn.hpp            — umbrella header
+src/
+  network.cpp        — non-template Network implementation
+tests/
+  test_fixed_string.cpp
+  test_traits.cpp
+  test_channel.cpp
+  test_node.cpp
+  test_network.cpp
+python/
+  kpn_python.cpp     — nanobind module entry point
+examples/
+  01_hello_pipeline/ … 09_opencv_cellshade/
+scripts/
+  render_readme.py   — regenerates README.md from README.md.in
+```

benchmarks/CMakeLists.txt

@@ -0,0 +1,14 @@
+cmake_minimum_required(VERSION 3.21)
+
+add_executable(bench_pipeline bench_pipeline.cpp)
+target_link_libraries(bench_pipeline PRIVATE kpn)
+target_compile_options(bench_pipeline PRIVATE -O3 -march=native)
+
+find_package(TBB QUIET)
+if(TBB_FOUND)
+    target_link_libraries(bench_pipeline PRIVATE TBB::tbb)
+    target_compile_definitions(bench_pipeline PRIVATE KPN_BENCH_TBB=1)
+    message(STATUS "TBB found — enabling TBB benchmarks")
+else()
+    message(STATUS "TBB not found — TBB benchmarks disabled")
+endif()

benchmarks/bench_pipeline.cpp

@@ -0,0 +1,526 @@
+// Throughput benchmark: items/second vs. graph topology and size.
+//
+// Topologies:
+//   chain   — linear depth D: push → n[0..D-1] → pop
+//   wide    — fanout<W>:      push → fanout → W parallel nodes → W pops
+//   diamond — push → fanout<2> → 2×2 nodes → 2 pops
+//
+// Two scheduling modes for each topology:
+//   private — each node owns a private ThreadPool(1)  [Node<>]
+//   pool    — all nodes share one ThreadPool(T)        [PoolNode<> + shared pool]
+//
+// Usage: ./bench_pipeline | tee results.csv
+
+#include <kpn/kpn.hpp>
+
+#ifdef KPN_BENCH_TBB
+#include <oneapi/tbb/flow_graph.h>
+namespace tbb_flow = oneapi::tbb::flow;
+#endif
+
+#include <array>
+#include <atomic>
+#include <chrono>
+#include <cstdio>
+#include <memory>
+#include <string>
+#include <thread>
+#include <vector>
+
+using namespace kpn;
+using namespace std::chrono_literals;
+using sclock = std::chrono::steady_clock;
+
+// ── configurable work ─────────────────────────────────────────────────────────
+
+static std::atomic<int> g_work_us{0};
+
+static int chain_fn(int x) {
+    int us = g_work_us.load(std::memory_order_relaxed);
+    if (us > 0) {
+        auto end = sclock::now() + std::chrono::microseconds(us);
+        while (sclock::now() < end);
+    }
+    return x;
+}
+
+using ChainNode     = Node<chain_fn, in<>, out<>>;
+using PoolChainNode = PoolNode<chain_fn, in<>, out<>>;
+
+// ── push helper: yield-spin on overflow (no artificial sleep latency) ─────────
+
+static void push_retry(Channel<int>& ch, int val) {
+    while (true) {
+        try   { ch.push(val); return; }
+        catch (const ChannelOverflowError&) { std::this_thread::yield(); }
+        catch (const ChannelClosedError&)   { return; }
+    }
+}
+
+// ── result ────────────────────────────────────────────────────────────────────
+
+struct Result {
+    const char* topology;
+    int         size;
+    int         work_us;
+    int         threads;     // 0 = private (1 thread per node), N = shared pool size
+    double      items_per_sec;
+    double      overhead_us;
+};
+
+// ── chain ─────────────────────────────────────────────────────────────────────
+
+static int items_for(int work_us, int depth = 1) {
+    int effective = std::max(1, work_us) * std::max(1, depth);
+    if (effective <= 1)     return 5000;
+    if (effective <= 10)    return 3000;
+    if (effective <= 100)   return 1000;
+    if (effective <= 1000)  return 200;
+    return 50;
+}
+
+static Result bench_chain(int depth, int work_us) {
+    const int N   = items_for(work_us, depth);
+    const int CAP = N;
+
+    std::vector<std::shared_ptr<Channel<int>>> chs;
+    for (int i = 0; i <= depth; ++i)
+        chs.push_back(std::make_shared<Channel<int>>(CAP));
+
+    std::vector<std::unique_ptr<ChainNode>> nodes;
+    for (int i = 0; i < depth; ++i) {
+        nodes.push_back(std::make_unique<ChainNode>(CAP));
+        nodes.back()->set_input_channel<0>(chs[i]);
+        nodes.back()->set_output_channel<0>(chs[i + 1].get());
+    }
+
+    for (auto& n : nodes) n->start();
+
+    std::atomic<sclock::time_point> t1;
+    std::thread reader([&] {
+        for (int i = 0; i < N; ++i) chs.back()->pop();
+        t1.store(sclock::now(), std::memory_order_release);
+    });
+
+    auto t0 = sclock::now();
+    std::thread pusher([&] {
+        for (int i = 0; i < N; ++i) push_retry(*chs[0], i);
+    });
+
+    pusher.join();
+    reader.join();
+    for (auto& n : nodes) n->stop();
+
+    double elapsed = std::chrono::duration<double>(
+        t1.load(std::memory_order_acquire) - t0).count();
+    // Subtract theoretical pipeline fill cost (depth-1)*W so that overhead
+    // reflects only framework latency, not the expected pipeline startup time.
+    double pipeline_us = static_cast<double>(work_us) * (N + depth - 1);
+    double wus = (elapsed * 1e6 - pipeline_us) / N;
+    return {"chain", depth, work_us, 0, N / elapsed, wus};
+}
+
+static Result bench_chain_pool(int depth, int work_us, int pool_threads) {
+    const int N   = items_for(work_us, depth);
+    const int CAP = N;
+
+    auto pool = std::make_shared<ThreadPool>(pool_threads);
+
+    std::vector<std::shared_ptr<Channel<int>>> chs;
+    for (int i = 0; i <= depth; ++i)
+        chs.push_back(std::make_shared<Channel<int>>(CAP));
+
+    std::vector<std::unique_ptr<PoolChainNode>> nodes;
+    for (int i = 0; i < depth; ++i) {
+        nodes.push_back(std::make_unique<PoolChainNode>(pool, CAP));
+        nodes.back()->set_input_channel<0>(chs[i]);
+        nodes.back()->set_output_channel<0>(chs[i + 1].get());
+    }
+
+    pool->start();
+    for (auto& n : nodes) n->start();
+
+    std::atomic<sclock::time_point> t1;
+    std::thread reader([&] {
+        for (int i = 0; i < N; ++i) chs.back()->pop();
+        t1.store(sclock::now(), std::memory_order_release);
+    });
+
+    auto t0 = sclock::now();
+    std::thread pusher([&] {
+        for (int i = 0; i < N; ++i) push_retry(*chs[0], i);
+    });
+
+    pusher.join();
+    reader.join();
+    for (auto& n : nodes) n->stop();
+    pool->stop();
+
+    double elapsed = std::chrono::duration<double>(
+        t1.load(std::memory_order_acquire) - t0).count();
+    double pipeline_us = static_cast<double>(work_us) * (N + depth - 1);
+    double wus = (elapsed * 1e6 - pipeline_us) / N;
+    return {"chain", depth, work_us, pool_threads, N / elapsed, wus};
+}
+
+// ── wide (fanout<W>) ──────────────────────────────────────────────────────────
+
+template<std::size_t W>
+static Result bench_wide(int work_us) {
+    const int N   = items_for(work_us);
+    const int CAP = N;
+
+    auto src_ch = std::make_shared<Channel<int>>(CAP);
+    auto fan    = std::make_unique<FanoutNode<int, W>>(CAP);
+    fan->template set_input_channel<0>(src_ch);
+
+    std::array<std::unique_ptr<ChainNode>,    W> nodes;
+    std::array<std::shared_ptr<Channel<int>>, W> sink_chs;
+
+    for (std::size_t i = 0; i < W; ++i) {
+        nodes[i]    = std::make_unique<ChainNode>(CAP);
+        sink_chs[i] = std::make_shared<Channel<int>>(CAP);
+        nodes[i]->template set_output_channel<0>(sink_chs[i].get());
+    }
+
+    [&]<std::size_t... Is>(std::index_sequence<Is...>) {
+        (fan->template set_output_channel<Is>(
+             &nodes[Is]->template input_channel<0>()), ...);
+    }(std::make_index_sequence<W>{});
+
+    fan->start();
+    for (auto& n : nodes) n->start();
+
+    std::array<std::thread, W> readers;
+    std::atomic<sclock::time_point> t1;
+    std::atomic<int> readers_done{0};
+
+    for (std::size_t w = 0; w < W; ++w) {
+        readers[w] = std::thread([&, w] {
+            for (int i = 0; i < N; ++i) sink_chs[w]->pop();
+            if (readers_done.fetch_add(1, std::memory_order_acq_rel) + 1
+                    == static_cast<int>(W))
+                t1.store(sclock::now(), std::memory_order_release);
+        });
+    }
+
+    auto t0 = sclock::now();
+    std::thread pusher([&] {
+        for (int i = 0; i < N; ++i) push_retry(*src_ch, i);
+    });
+
+    pusher.join();
+    for (auto& r : readers) r.join();
+    fan->stop();
+    for (auto& n : nodes) n->stop();
+
+    double elapsed = std::chrono::duration<double>(
+        t1.load(std::memory_order_acquire) - t0).count();
+    double wus = (elapsed * 1e6) / N - static_cast<double>(work_us);
+    return {"wide", static_cast<int>(W), work_us, 0, N / elapsed, wus};
+}
+
+template<std::size_t W>
+static Result bench_wide_pool(int work_us, int pool_threads) {
+    const int N   = items_for(work_us);
+    const int CAP = N;
+
+    auto pool   = std::make_shared<ThreadPool>(pool_threads);
+    auto src_ch = std::make_shared<Channel<int>>(CAP);
+    auto fan    = std::make_unique<FanoutNode<int, W>>(CAP);
+    fan->template set_input_channel<0>(src_ch);
+
+    std::array<std::unique_ptr<PoolChainNode>, W> nodes;
+    std::array<std::shared_ptr<Channel<int>>,  W> sink_chs;
+
+    for (std::size_t i = 0; i < W; ++i) {
+        nodes[i]    = std::make_unique<PoolChainNode>(pool, CAP);
+        sink_chs[i] = std::make_shared<Channel<int>>(CAP);
+        nodes[i]->template set_output_channel<0>(sink_chs[i].get());
+    }
+
+    [&]<std::size_t... Is>(std::index_sequence<Is...>) {
+        (fan->template set_output_channel<Is>(
+             &nodes[Is]->template input_channel<0>()), ...);
+    }(std::make_index_sequence<W>{});
+
+    fan->start();
+    pool->start();
+    for (auto& n : nodes) n->start();
+
+    std::array<std::thread, W> readers;
+    std::atomic<sclock::time_point> t1;
+    std::atomic<int> readers_done{0};
+
+    for (std::size_t w = 0; w < W; ++w) {
+        readers[w] = std::thread([&, w] {
+            for (int i = 0; i < N; ++i) sink_chs[w]->pop();
+            if (readers_done.fetch_add(1, std::memory_order_acq_rel) + 1
+                    == static_cast<int>(W))
+                t1.store(sclock::now(), std::memory_order_release);
+        });
+    }
+
+    auto t0 = sclock::now();
+    std::thread pusher([&] {
+        for (int i = 0; i < N; ++i) push_retry(*src_ch, i);
+    });
+
+    pusher.join();
+    for (auto& r : readers) r.join();
+    fan->stop();
+    for (auto& n : nodes) n->stop();
+    pool->stop();
+
+    double elapsed = std::chrono::duration<double>(
+        t1.load(std::memory_order_acquire) - t0).count();
+    double wus = (elapsed * 1e6) / N - static_cast<double>(work_us);
+    return {"wide", static_cast<int>(W), work_us, pool_threads, N / elapsed, wus};
+}
+
+// ── diamond ───────────────────────────────────────────────────────────────────
+
+static Result bench_diamond(int work_us) {
+    const int N   = items_for(work_us, 2);
+    const int CAP = N;
+
+    auto src_ch = std::make_shared<Channel<int>>(CAP);
+    auto fan    = std::make_unique<FanoutNode<int, 2>>(CAP);
+    fan->template set_input_channel<0>(src_ch);
+
+    auto nL  = std::make_unique<ChainNode>(CAP);
+    auto nR  = std::make_unique<ChainNode>(CAP);
+    auto nL2 = std::make_unique<ChainNode>(CAP);
+    auto nR2 = std::make_unique<ChainNode>(CAP);
+    auto chL  = std::make_shared<Channel<int>>(CAP);
+    auto chR  = std::make_shared<Channel<int>>(CAP);
+    auto snkL = std::make_shared<Channel<int>>(CAP);
+    auto snkR = std::make_shared<Channel<int>>(CAP);
+
+    fan->template set_output_channel<0>(&nL->template input_channel<0>());
+    fan->template set_output_channel<1>(&nR->template input_channel<0>());
+    nL->set_output_channel<0>(chL.get());
+    nR->set_output_channel<0>(chR.get());
+    nL2->set_input_channel<0>(chL);
+    nR2->set_input_channel<0>(chR);
+    nL2->set_output_channel<0>(snkL.get());
+    nR2->set_output_channel<0>(snkR.get());
+
+    fan->start(); nL->start(); nR->start(); nL2->start(); nR2->start();
+
+    std::atomic<sclock::time_point> t1;
+    std::atomic<int> done{0};
+    auto make_reader = [&](Channel<int>& ch) {
+        return std::thread([&] {
+            for (int i = 0; i < N; ++i) ch.pop();
+            if (done.fetch_add(1, std::memory_order_acq_rel) + 1 == 2)
+                t1.store(sclock::now(), std::memory_order_release);
+        });
+    };
+    auto rL = make_reader(*snkL);
+    auto rR = make_reader(*snkR);
+
+    auto t0 = sclock::now();
+    std::thread pusher([&] {
+        for (int i = 0; i < N; ++i) push_retry(*src_ch, i);
+    });
+
+    pusher.join(); rL.join(); rR.join();
+    fan->stop(); nL->stop(); nR->stop(); nL2->stop(); nR2->stop();
+
+    double elapsed = std::chrono::duration<double>(
+        t1.load(std::memory_order_acquire) - t0).count();
+    double wus = (elapsed * 1e6) / N - static_cast<double>(work_us);
+    return {"diamond", 4, work_us, 0, N / elapsed, wus};
+}
+
+static Result bench_diamond_pool(int work_us, int pool_threads) {
+    const int N   = items_for(work_us, 2);
+    const int CAP = N;
+
+    auto pool   = std::make_shared<ThreadPool>(pool_threads);
+    auto src_ch = std::make_shared<Channel<int>>(CAP);
+    auto fan    = std::make_unique<FanoutNode<int, 2>>(CAP);
+    fan->template set_input_channel<0>(src_ch);
+
+    auto nL  = std::make_unique<PoolChainNode>(pool, CAP);
+    auto nR  = std::make_unique<PoolChainNode>(pool, CAP);
+    auto nL2 = std::make_unique<PoolChainNode>(pool, CAP);
+    auto nR2 = std::make_unique<PoolChainNode>(pool, CAP);
+    auto chL  = std::make_shared<Channel<int>>(CAP);
+    auto chR  = std::make_shared<Channel<int>>(CAP);
+    auto snkL = std::make_shared<Channel<int>>(CAP);
+    auto snkR = std::make_shared<Channel<int>>(CAP);
+
+    fan->template set_output_channel<0>(&nL->template input_channel<0>());
+    fan->template set_output_channel<1>(&nR->template input_channel<0>());
+    nL->set_output_channel<0>(chL.get());
+    nR->set_output_channel<0>(chR.get());
+    nL2->set_input_channel<0>(chL);
+    nR2->set_input_channel<0>(chR);
+    nL2->set_output_channel<0>(snkL.get());
+    nR2->set_output_channel<0>(snkR.get());
+
+    fan->start();
+    pool->start();
+    nL->start(); nR->start(); nL2->start(); nR2->start();
+
+    std::atomic<sclock::time_point> t1;
+    std::atomic<int> done{0};
+    auto make_reader = [&](Channel<int>& ch) {
+        return std::thread([&] {
+            for (int i = 0; i < N; ++i) ch.pop();
+            if (done.fetch_add(1, std::memory_order_acq_rel) + 1 == 2)
+                t1.store(sclock::now(), std::memory_order_release);
+        });
+    };
+    auto rL = make_reader(*snkL);
+    auto rR = make_reader(*snkR);
+
+    auto t0 = sclock::now();
+    std::thread pusher([&] {
+        for (int i = 0; i < N; ++i) push_retry(*src_ch, i);
+    });
+
+    pusher.join(); rL.join(); rR.join();
+    fan->stop();
+    nL->stop(); nR->stop(); nL2->stop(); nR2->stop();
+    pool->stop();
+
+    double elapsed = std::chrono::duration<double>(
+        t1.load(std::memory_order_acquire) - t0).count();
+    double wus = (elapsed * 1e6) / N - static_cast<double>(work_us);
+    return {"diamond", 4, work_us, pool_threads, N / elapsed, wus};
+}
+
+// ── TBB flow graph ────────────────────────────────────────────────────────────
+#ifdef KPN_BENCH_TBB
+
+static Result bench_chain_tbb(int depth, int work_us) {
+    const int N = items_for(work_us, depth);
+
+    tbb_flow::graph g;
+    using FN = tbb_flow::function_node<int, int>;
+    std::vector<std::unique_ptr<FN>> nodes;
+    nodes.reserve(depth);
+    for (int i = 0; i < depth; ++i)
+        nodes.push_back(std::make_unique<FN>(g, tbb_flow::serial,
+            [](int x) -> int { return chain_fn(x); }));
+    for (int i = 0; i + 1 < depth; ++i)
+        tbb_flow::make_edge(*nodes[i], *nodes[i + 1]);
+
+    auto t0 = sclock::now();
+    for (int i = 0; i < N; ++i) nodes[0]->try_put(i);
+    g.wait_for_all();
+    auto t1 = sclock::now();
+
+    double elapsed    = std::chrono::duration<double>(t1 - t0).count();
+    double pipeline_us = static_cast<double>(work_us) * (N + depth - 1);
+    double wus        = (elapsed * 1e6 - pipeline_us) / N;
+    return {"chain_tbb", depth, work_us, -1, N / elapsed, wus};
+}
+
+template<std::size_t W>
+static Result bench_wide_tbb(int work_us) {
+    const int N = items_for(work_us);
+
+    tbb_flow::graph g;
+    tbb_flow::broadcast_node<int> fan(g);
+    using FN = tbb_flow::function_node<int, int>;
+    std::array<std::unique_ptr<FN>, W> nodes;
+    for (auto& n : nodes) {
+        n = std::make_unique<FN>(g, tbb_flow::serial,
+            [](int x) -> int { return chain_fn(x); });
+        tbb_flow::make_edge(fan, *n);
+    }
+
+    auto t0 = sclock::now();
+    for (int i = 0; i < N; ++i) fan.try_put(i);
+    g.wait_for_all();
+    auto t1 = sclock::now();
+
+    double elapsed = std::chrono::duration<double>(t1 - t0).count();
+    double wus     = (elapsed * 1e6) / N - static_cast<double>(work_us);
+    return {"wide_tbb", static_cast<int>(W), work_us, -1, N / elapsed, wus};
+}
+
+static Result bench_diamond_tbb(int work_us) {
+    const int N = items_for(work_us, 2);
+
+    tbb_flow::graph g;
+    tbb_flow::broadcast_node<int> fan(g);
+    using FN = tbb_flow::function_node<int, int>;
+    auto fn = [](int x) -> int { return chain_fn(x); };
+    FN nL(g, tbb_flow::serial, fn), nR(g, tbb_flow::serial, fn);
+    FN nL2(g, tbb_flow::serial, fn), nR2(g, tbb_flow::serial, fn);
+    tbb_flow::make_edge(fan, nL);  tbb_flow::make_edge(fan, nR);
+    tbb_flow::make_edge(nL, nL2);  tbb_flow::make_edge(nR, nR2);
+
+    auto t0 = sclock::now();
+    for (int i = 0; i < N; ++i) fan.try_put(i);
+    g.wait_for_all();
+    auto t1 = sclock::now();
+
+    double elapsed = std::chrono::duration<double>(t1 - t0).count();
+    double wus     = (elapsed * 1e6) / N - static_cast<double>(work_us);
+    return {"diamond_tbb", 4, work_us, -1, N / elapsed, wus};
+}
+#endif // KPN_BENCH_TBB
+
+// ── main ──────────────────────────────────────────────────────────────────────
+
+int main() {
+    const int work_amts[]  = {10, 100, 1000};
+    const int pool_sizes[] = {1, 2, 4};
+
+    std::fprintf(stderr, "%-12s %-8s %-10s %-8s %-18s %-20s\n",
+                 "topology", "size", "work_us", "threads", "items/sec", "overhead_us/item");
+    std::fprintf(stderr, "%s\n", std::string(78, '-').c_str());
+    std::printf("topology,size,work_us,threads,items_per_sec,overhead_us_per_item\n");
+
+    auto emit = [](const Result& r) {
+        std::string sched = r.threads  < 0 ? "tbb"
+                          : r.threads == 0 ? "priv"
+                          : std::to_string(r.threads);
+        std::fprintf(stderr, "%-12s %-8d %-10d %-8s %-18.0f %-20.1f\n",
+                     r.topology, r.size, r.work_us, sched.c_str(),
+                     r.items_per_sec, r.overhead_us);
+        std::printf("%s,%d,%d,%s,%.0f,%.2f\n",
+                    r.topology, r.size, r.work_us, sched.c_str(),
+                    r.items_per_sec, r.overhead_us);
+        std::fflush(stdout);
+    };
+
+    for (int w : work_amts) {
+        g_work_us.store(w, std::memory_order_relaxed);
+        std::fprintf(stderr, "\n── work_us=%-4d private pools ───────────────────────────────────────\n", w);
+
+        for (int d : {1, 2, 4, 8, 16, 32}) emit(bench_chain(d, w));
+        emit(bench_wide<1>(w));
+        emit(bench_wide<2>(w));
+        emit(bench_wide<3>(w));
+        emit(bench_wide<4>(w));
+        emit(bench_diamond(w));
+
+        for (int pt : pool_sizes) {
+            std::fprintf(stderr, "\n── work_us=%-4d shared pool (%d thread%s) ─────────────────────────────\n",
+                         w, pt, pt == 1 ? "" : "s");
+            for (int d : {1, 2, 4, 8, 16, 32}) emit(bench_chain_pool(d, w, pt));
+            emit(bench_wide_pool<1>(w, pt));
+            emit(bench_wide_pool<2>(w, pt));
+            emit(bench_wide_pool<3>(w, pt));
+            emit(bench_wide_pool<4>(w, pt));
+            emit(bench_diamond_pool(w, pt));
+        }
+
+#ifdef KPN_BENCH_TBB
+        std::fprintf(stderr, "\n── work_us=%-4d TBB flow graph ──────────────────────────────────────\n", w);
+        for (int d : {1, 2, 4, 8, 16, 32}) emit(bench_chain_tbb(d, w));
+        emit(bench_wide_tbb<1>(w));
+        emit(bench_wide_tbb<2>(w));
+        emit(bench_wide_tbb<3>(w));
+        emit(bench_wide_tbb<4>(w));
+        emit(bench_diamond_tbb(w));
+#endif
+    }
+}

docs/channels.md

@@ -0,0 +1,55 @@
+# Channels
+
+A `Channel<T>` is a lock-free SPSC (single-producer, single-consumer) ring buffer with atomic wait/notify.
+
+## Semantics
+
+- **Bounded**: fixed capacity set at construction. Default is 5 items.
+- **Backpressure**: when full, `push()` throws `ChannelOverflowError` immediately — no blocking, no spin.
+- **Blocking consumer**: `pop()` blocks until an item is available or the channel is disabled.
+- **Disable**: `channel.disable()` stops accepting pushes and unblocks any waiting `pop()` with `ChannelClosedError`.
+
+## Storage policy
+
+Small trivially-copyable types (≤ 8 bytes) are stored by value. Larger types are heap-allocated and passed via `shared_ptr<const T>` — one allocation per push, zero-copy fan-out:
+
+```cpp
+--8<-- "examples/04_storage_policy/main.cpp:storage_policy_spec"
+```
+
+Specialize `kpn::ChannelDataSize<T>` for accurate bandwidth reporting on heap-owning types:
+
+```cpp
+template<>
+struct kpn::ChannelDataSize<cv::Mat> {
+    static std::size_t bytes(const cv::Mat& m) { return m.total() * m.elemSize(); }
+};
+```
+
+## Named ports
+
+`in<"name">` and `out<"name">` tag nodes for readable wiring:
+
+```cpp
+--8<-- "examples/02_named_ports/main.cpp:named_port_creation"
+```
+
+Named ports are checked at compile time — a typo in a port name is a compile error.
+
+## Capacity tuning
+
+Set capacity per node at construction:
+
+```cpp
+auto node = make_node<my_func>(/*capacity=*/20);
+```
+
+Capacity is rounded up internally to the next power of two. Monitor fill levels via diagnostics to tune for your workload — a too-small capacity causes overflows; a too-large one wastes memory and hides producer/consumer speed mismatches.
+
+## Spin count
+
+`Channel` spins for up to ~4 µs (200 `pause` hints at ~20 ns each on x86) before sleeping on a futex. Set to 0 for power-constrained or predominantly-idle pipelines:
+
+```cpp
+Channel<int> ch(/*capacity=*/5, /*spin_count=*/0);
+```

docs/error-handling.md

@@ -0,0 +1,84 @@
+# Error Handling & Events
+
+KPN++ provides three complementary layers for observing and reacting to failures.
+
+---
+
+## 1. Per-node error handler
+
+Called when a node's function throws an unhandled exception. Return `true` to skip the failed invocation and keep running; `false` to stop the node.
+
+```cpp
+--8<-- "examples/15_node_error_handler/main.cpp:error_handler"
+```
+
+When a node stops (either from `false` return or no handler installed), it:
+
+1. Disables its **input** channels — upstream stops pushing into dead queues.
+2. Disables its **output** channels — downstream nodes receive `ChannelClosedError` on their next pop, propagating the shutdown naturally through the graph.
+
+---
+
+## 2. Per-node overflow callback
+
+Fired with a timestamp each time an output push is dropped because the channel is full. The node name is known at registration so it is not included — keeping the callback zero-overhead when unused.
+
+```cpp
+--8<-- "examples/16_event_callbacks/main.cpp:per_node_callback"
+```
+
+!!! note
+    The callback is purely informational — the node always continues after an overflow. To stop the node on overflow, call `node.stop()` from inside the callback.
+
+A matching `set_closed_callback()` fires (also with just a timestamp) when the node stops due to a closed upstream channel:
+
+```cpp
+node.set_closed_callback([](std::chrono::steady_clock::time_point ts) {
+    std::cerr << "node stopped at t=" << ts.time_since_epoch().count() << '\n';
+});
+```
+
+Each node holds two callback slots per event type — one user-set (registered above) and one injected by the network (see below). Both fire independently.
+
+---
+
+## 3. Network-level event handler
+
+One callback for the whole network. Receives the node name (captured in a closure by the network at `build()` / `start()`), a `NodeEvent`, and a timestamp:
+
+```cpp
+--8<-- "examples/16_event_callbacks/main.cpp:network_event_handler"
+```
+
+`NodeEvent` values:
+
+| Value | Meaning |
+|---|---|
+| `NodeEvent::Overflow` | An output push was dropped (channel full) |
+| `NodeEvent::Closed` | The node stopped (crash or upstream close cascade) |
+
+The network handler and any per-node callbacks are **independent** — both fire when set.
+
+---
+
+## Complete example
+
+`examples/16_event_callbacks/main.cpp` shows a fast producer overflowing a slow consumer, with both a per-node overflow callback and a network-level event handler active simultaneously.
+
+Node functions:
+
+```cpp
+--8<-- "examples/16_event_callbacks/main.cpp:node_fns"
+```
+
+Per-node overflow callback:
+
+```cpp
+--8<-- "examples/16_event_callbacks/main.cpp:per_node_callback"
+```
+
+Network-level event handler:
+
+```cpp
+--8<-- "examples/16_event_callbacks/main.cpp:network_event_handler"
+```

docs/examples.md

@@ -0,0 +1,40 @@
+# Examples
+
+All C++ examples are built by default and registered as CTest smoke tests. Run them all with:
+
+```bash
+ctest --test-dir build -L examples
+```
+
+## Index
+
+| Example | What it shows |
+|---|---|
+| `01_hello_pipeline` | Linear pipeline, index-based port wiring |
+| `02_named_ports` | `in<>`/`out<>` name tags, named port access |
+| `03_multi_output` | Tuple-returning node, per-element routing |
+| `04_storage_policy` | `channel_storage_policy` specialisation |
+| `05_error_handling` | Diagnostics handler, overflow channel stats |
+| `06_watchdog` | Watchdog interval, stall detection |
+| `10_static_hello_pipeline` | `StaticNetwork` + `make_network()` |
+| `11_static_fanout` | `StaticNetwork` with `FanoutNode` |
+| `15_node_error_handler` | `set_error_handler()` — skip or stop on exception |
+| `16_event_callbacks` | `set_overflow_callback()`, `set_event_handler()` |
+
+## OpenCV examples (optional)
+
+Built only when OpenCV ≥ 4 is found:
+
+| Example | What it shows |
+|---|---|
+| `09_opencv_cellshade` | Real-time cell-shading on webcam; `MainThreadNode` for display |
+| `12_static_cellshade` | Same pipeline as a `StaticNetwork` |
+| `13_debug_cellshade` | Web debug UI overlay on the cell-shading pipeline |
+
+Run the cell-shading example:
+
+```bash
+./build/examples/09_opencv_cellshade
+# Press 'q' or close the window to stop.
+# Falls back to an animated synthetic pattern if no webcam is found.
+```

docs/fanout.md

@@ -0,0 +1,44 @@
+# Fan-out & Routing
+
+## FanoutNode
+
+Reads one item and pushes a copy to each of N output channels. All downstream nodes receive every item.
+
+```cpp
+auto fan = make_fanout<Image, 2>(/*capacity=*/8);
+
+net.connect("src",  src.output<0>(),  "fan",   fan.input<0>())
+   .connect("fan",  fan.output<0>(),  "nodeA", nodeA.input<0>())
+   .connect("fan",  fan.output<1>(),  "nodeB", nodeB.input<0>());
+```
+
+If one downstream channel overflows, that output drops the item independently — the other outputs are unaffected.
+
+See `examples/11_static_fanout`.
+
+## RouterNode
+
+Reads one item and pushes it to exactly one of N outputs, chosen by a selector function:
+
+```cpp
+auto router = make_router<Frame, 3>(
+    [](const Frame& f) -> std::size_t { return f.stream_id % 3; });
+
+net.connect("src",    src.output<0>(),     "router", router.input<0>())
+   .connect("router", router.output<0>(),  "nodeA",  nodeA.input<0>())
+   .connect("router", router.output<1>(),  "nodeB",  nodeB.input<0>())
+   .connect("router", router.output<2>(),  "nodeC",  nodeC.input<0>());
+```
+
+If the selector returns `>= N` the item is silently dropped.
+
+## FilterNode
+
+Reads one item and passes it downstream only when a predicate returns `true`:
+
+```cpp
+auto filt = make_filter<Frame>([](const Frame& f) { return f.valid; });
+
+net.connect("src",  src.output<0>(),  "filt", filt.input<0>())
+   .connect("filt", filt.output<0>(), "dst",  dst.input<0>());
+```

docs/getting-started.md

@@ -0,0 +1,76 @@
+# Getting Started
+
+## Requirements
+
+| Dependency | Version | Notes |
+|---|---|---|
+| CMake | ≥ 3.21 | |
+| C++ compiler | GCC ≥ 11, Clang ≥ 13 | C++20 required |
+| nanobind | ≥ 2.1 | auto-fetched; Python ≥ 3.8 |
+| Catch2 | v3 | auto-fetched for tests |
+| OpenCV | ≥ 4 | optional; only for examples 09/12/13 |
+
+## Build
+
+```bash
+cmake -B build                          # core + tests + C++ examples
+cmake --build build --parallel
+ctest --test-dir build                  # run all tests including example smoke tests
+```
+
+Enable Python bindings:
+
+```bash
+cmake -B build -DKPN_BUILD_PYTHON=ON
+cmake --build build --parallel
+```
+
+Skip examples:
+
+```bash
+cmake -B build -DKPN_BUILD_EXAMPLES=OFF
+```
+
+## Your first pipeline
+
+Three functions — source, transform, sink — wired into a `Network`:
+
+```cpp
+--8<-- "examples/01_hello_pipeline/main.cpp:basic_node_fns"
+```
+
+Create nodes, connect them, build and run:
+
+```cpp
+--8<-- "examples/01_hello_pipeline/main.cpp:network_build"
+```
+
+That's it. Types are inferred from function signatures. The channel between `src` and `dbl` carries `int`; the channel between `dbl` and `prn` also carries `int`. A type mismatch is a compile error.
+
+## Named ports
+
+For nodes with multiple inputs or outputs, name the ports for clarity:
+
+```cpp
+--8<-- "examples/02_named_ports/main.cpp:named_port_creation"
+```
+
+Wire by name instead of index:
+
+```cpp
+--8<-- "examples/02_named_ports/main.cpp:named_port_network"
+```
+
+## Multi-output nodes
+
+Return a `std::tuple` to fan out to multiple downstream nodes:
+
+```cpp
+--8<-- "examples/03_multi_output/main.cpp:multi_output_fn"
+```
+
+Wire each tuple element to its own downstream node:
+
+```cpp
+--8<-- "examples/03_multi_output/main.cpp:fanout_network"
+```

docs/index.md

@@ -0,0 +1,39 @@
+# KPN++
+
+A C++20 [Kahn Process Network](https://en.wikipedia.org/wiki/Kahn_process_networks) library. Each node wraps a plain function and runs concurrently, communicating with downstream nodes via bounded FIFO channels. Includes Python bindings via nanobind.
+
+---
+
+## Why KPN++?
+
+- **Zero boilerplate** — wrap any callable as a node; types flow automatically from the function signature
+- **Bounded channels** — backpressure is structural, not bolted on
+- **Observable** — per-node and network-level callbacks for overflow and stop events; diagnostics snapshots; optional web UI
+- **Composable** — `Network` for runtime wiring, `StaticNetwork` for compile-time topology with zero overhead
+
+---
+
+## Quick example
+
+```cpp
+#include <kpn/kpn.hpp>
+using namespace kpn;
+
+--8<-- "examples/01_hello_pipeline/main.cpp:basic_node_fns"
+
+int main() {
+--8<-- "examples/01_hello_pipeline/main.cpp:network_build"
+}
+```
+
+---
+
+## Install & build
+
+```bash
+cmake -B build
+cmake --build build --parallel
+ctest --test-dir build          # unit tests + example smoke tests
+```
+
+See [Getting Started](getting-started.md) for full build options.

docs/network.md

@@ -0,0 +1,67 @@
+# Networks
+
+A `Network` wires nodes together at runtime using a builder chain.
+
+## Building a network
+
+```cpp
+--8<-- "examples/01_hello_pipeline/main.cpp:network_build"
+```
+
+The builder chain:
+
+| Method | Purpose |
+|---|---|
+| `.add(name, node)` | Register a node; assigns its name |
+| `.connect(src, port, dst, port)` | Wire one output port to one input port |
+| `.build()` | Compute topological order; inject network callbacks |
+| `.start()` | Start nodes in topological order |
+| `.stop()` | Stop all nodes immediately |
+| `.shutdown()` | Graceful drain: stop sources first, wait for channels to empty, then stop downstream |
+
+## Port access
+
+Ports are accessed by index or by name:
+
+```cpp
+// By index
+net.connect("src", src.output<0>(), "dst", dst.input<0>());
+
+// By name (requires named ports)
+--8<-- "examples/02_named_ports/main.cpp:named_port_network"
+```
+
+## Diagnostics
+
+Install a diagnostics handler to receive periodic snapshots of every node and channel:
+
+```cpp
+--8<-- "examples/05_error_handling/main.cpp:diagnostics_handler"
+```
+
+Or print a full report at any time:
+
+```cpp
+net.print_diagnostics();   // writes to stderr by default
+net.print_diagnostics(std::cout);
+```
+
+## Network-level event handler
+
+Observe overflow and node-stop events across the entire network in one place:
+
+```cpp
+--8<-- "examples/16_event_callbacks/main.cpp:network_event_handler"
+```
+
+`NodeEvent` is either `NodeEvent::Overflow` (item dropped on full channel) or `NodeEvent::Closed` (node stopped due to crash or closed upstream channel). See [Error Handling & Events](error-handling.md).
+
+## Shutdown
+
+`net.stop()` halts immediately — all nodes stop in reverse topological order.
+
+`net.shutdown()` drains gracefully: source nodes stop first; their output channels are polled until empty; then the next layer stops, and so on. This ensures no items are lost if downstream nodes are still consuming.
+
+## StaticNetwork
+
+For zero-overhead compile-time topology, see [Static Networks](static-network.md).

docs/nodes.md

@@ -0,0 +1,81 @@
+# Nodes
+
+A node wraps any callable. Its input types are inferred from the function's parameter list; its output types from the return type.
+
+## Node types
+
+| Type | Thread model | Use case |
+|---|---|---|
+| `Node<Func>` | Dedicated thread per node | Default — simplest, most isolated |
+| `PoolNode<Func>` | Shared `ThreadPool` | Many nodes, resource-bounded execution |
+| `InterruptNode<Func>` | Event-driven, no thread | Camera frame ready, timer tick, socket |
+| `FanoutNode<T, N>` | Dedicated thread | Broadcast one item to N outputs |
+| `RouterNode<T, N>` | Dedicated thread | Route one item to one of N outputs |
+| `FilterNode<T>` | Dedicated thread | Pass items matching a predicate |
+
+## Creating nodes
+
+All node types are created via factory functions that infer types from the callable:
+
+```cpp
+// Free function — simplest case
+auto node = make_node<my_func>();
+
+// Stateful functor (operator() is the function)
+MyProcessor proc;
+auto node = make_node(proc);
+
+// Pool node — shares a ThreadPool with other nodes
+auto pool = std::make_shared<ThreadPool>(4);
+auto node = make_pool_node<my_func>(pool);
+
+// Interrupt node — triggered externally
+auto sched = std::make_shared<ThreadPool>(2);
+auto node  = make_interrupt_node<produce_frame>(sched, out<"frame">{});
+camera_sdk.on_frame_ready(node.get_trigger());
+```
+
+## Channel capacity
+
+Each node's input FIFO has a configurable capacity (default 5):
+
+```cpp
+auto node = make_node<my_func>(/*capacity=*/20);
+auto node = make_pool_node<my_func>(pool, /*capacity=*/20);
+```
+
+When an upstream push would exceed capacity, `ChannelOverflowError` is thrown and the item is dropped. See [Error Handling & Events](error-handling.md) to observe and react to this.
+
+## Source nodes
+
+A node with no inputs is a source. It self-submits immediately on `start()` and re-submits after each execution:
+
+```cpp
+static int produce() {
+    std::this_thread::sleep_for(std::chrono::milliseconds(10));
+    return ++counter;
+}
+auto src = make_node<produce>();
+```
+
+!!! tip
+    Source nodes must sleep or yield to avoid overflowing their output channel. The channel capacity provides the only bound.
+
+## Sink nodes
+
+A node with a `void` return is a sink — it consumes items without producing output:
+
+```cpp
+static void print_it(int x) { std::cout << x << '\n'; }
+auto snk = make_node<print_it>();
+```
+
+## Error handler
+
+When a node's function throws an unhandled exception, the default behaviour is to stop the node (disabling its channels so the shutdown cascades downstream). Install a handler to override:
+
+```cpp
+--8<-- "examples/15_node_error_handler/main.cpp:error_handler"
+```
+
+See [Error Handling & Events](error-handling.md) for the full picture.

docs/requirements.txt

@@ -0,0 +1,3 @@
+mkdocs>=1.5
+mkdocs-material>=9.5
+pymdown-extensions>=10.0

docs/shared-resource.md

@@ -0,0 +1,42 @@
+# Shared Resources
+
+`SharedResource<T>` arbitrates exclusive access to a resource (ONNX session, CUDA stream, serial port) across multiple nodes using a priority-based waiter queue with starvation prevention.
+
+## Usage
+
+```cpp
+#include <kpn/shared_resource.hpp>
+using namespace kpn;
+
+SharedResource<OnnxSession> model(session_args...);
+
+static cv::Mat run_inference(cv::Mat frame) {
+    // Acquires the model; releases automatically on scope exit.
+    auto guard = model.acquire_balanced(in_channel, out_channel);
+    return guard->Run(frame);
+}
+```
+
+## Acquire modes
+
+| Method | Priority |
+|---|---|
+| `acquire()` | Equal (fair FIFO) |
+| `acquire(fn)` | Custom — `fn()` returns `float` in `[0, 1]` |
+| `acquire_balanced(in_ch, out_ch)` | `input_fill × output_headroom` — highest urgency wins |
+
+`acquire_balanced` favours nodes with full input queues and empty output queues — the node that has the most work to do and nowhere to stall wins the resource next.
+
+## Starvation prevention
+
+Each waiter's effective score grows with elapsed wait time (`0.05` per second by default), ensuring a low-priority node eventually gets served regardless of how frequently higher-priority nodes compete.
+
+## Diagnostics
+
+Register with the network for snapshot reporting:
+
+```cpp
+net.register_resource("model", &model);
+```
+
+The diagnostics table then shows acquisition count, mean wait time, and current waiter count.

docs/static-network.md

@@ -0,0 +1,46 @@
+# Static Networks
+
+`StaticNetwork` encodes the entire topology at compile time using a `make_network()` builder. Nodes and channel types are verified statically with zero runtime overhead.
+
+## Usage
+
+```cpp
+#include <kpn/kpn.hpp>
+using namespace kpn;
+
+static int  produce()        { return 42; }
+static int  double_it(int x) { return x * 2; }
+static void print_it(int x)  { std::cout << x << '\n'; }
+
+int main() {
+    auto src = make_node<produce>  ();
+    auto dbl = make_node<double_it>();
+    auto prn = make_node<print_it> ();
+
+    auto net = make_network(
+        edge(src, src.output<0>(), dbl, dbl.input<0>()),
+        edge(dbl, dbl.output<0>(), prn, prn.input<0>())
+    );
+
+    net.set_event_handler([](std::string_view name, NodeEvent ev, auto ts) {
+        // same API as Network
+    });
+
+    net.start();
+    std::this_thread::sleep_for(std::chrono::milliseconds(100));
+    net.stop();
+}
+```
+
+See `examples/10_static_hello_pipeline` and `examples/11_static_fanout`.
+
+## When to use
+
+| | `Network` | `StaticNetwork` |
+|---|---|---|
+| Topology known at | Runtime | Compile time |
+| Type checking | Runtime (`dynamic_cast`) | Compile time |
+| Overhead | Minimal | Zero |
+| Flexibility | Add nodes dynamically | Fixed at compile time |
+
+For most applications `Network` is sufficient. Use `StaticNetwork` when you need the absolute minimum overhead or want compile-time topology verification.

examples/01_hello_pipeline/main.cpp

@@ -7,16 +7,20 @@
 //
 //  [produce] --int--> [double_it] --int--> [print_it]
 
+// --8<-- [start:basic_node_fns]
 static int  produce()        { return 42; }
 static int  double_it(int x) { return x * 2; }
 static void print_it(int x)  { std::cout << "result: " << x << '\n'; }
+// --8<-- [end:basic_node_fns]
 
 int main() {
     using namespace kpn;
 
+    // --8<-- [start:index_only_nodes]
     auto src  = make_node<produce>(5);
     auto dbl  = make_node<double_it>(5);
     auto sink = make_node<print_it>(5);
+    // --8<-- [end:index_only_nodes]
 
     // Wire channels
     auto& dbl_in  = dbl.input_channel<0>();
@@ -24,6 +28,7 @@ int main() {
     src.set_output_channel<0>(&dbl_in);
     dbl.set_output_channel<0>(&sink_in);
 
+    // --8<-- [start:network_build]
     Network net;
     net.add("src",  src)
        .add("dbl",  dbl)
@@ -35,4 +40,5 @@ int main() {
     net.start();
     std::this_thread::sleep_for(std::chrono::milliseconds(100));
     net.stop();
+    // --8<-- [end:network_build]
 }

examples/02_named_ports/main.cpp

@@ -54,17 +54,18 @@ static void report(int count, std::vector<std::string> words) {
 int main() {
     using namespace kpn;
 
+    // --8<-- [start:named_port_creation]
     // tokenise: no inputs, one named output "words"
     auto tok = make_node<tokenise>(out<"words">{}, 4);
 
     // count_words: named input "words", named outputs "count" and "words"
     auto cnt = make_node<count_words>(in<"words">{}, out<"count", "words">{}, 4);
 
-    // report: two named inputs — note the function takes (int, vector<string>)
-    // so we need two separate input ports wired independently
-    // For a two-input sink we wire each output of cnt to a different input of report
+    // report: two named inputs
     auto snk = make_node<report>(in<"count", "words">{}, 4);
+    // --8<-- [end:named_port_creation]
 
+    // --8<-- [start:named_port_network]
     Network net;
     net.add("tok", tok)
        .add("cnt", cnt)
@@ -77,4 +78,5 @@ int main() {
     net.start();
     std::this_thread::sleep_for(std::chrono::milliseconds(500));
     net.stop();
+    // --8<-- [end:named_port_network]
 }

examples/03_multi_output/main.cpp

@@ -33,6 +33,7 @@ static std::string generate() {
     return pairs[gen_index++ % 5];
 }
 
+// --8<-- [start:multi_output_fn]
 // Multi-output: returns (key, value) as a tuple — KPN++ routes each element
 // to its own output port automatically.
 static std::tuple<std::string, std::string> parse(std::string kv) {
@@ -40,6 +41,7 @@ static std::tuple<std::string, std::string> parse(std::string kv) {
     if (sep == std::string::npos) return {kv, ""};
     return {kv.substr(0, sep), kv.substr(sep + 1)};
 }
+// --8<-- [end:multi_output_fn]
 
 static void print_key(std::string key) {
     std::cout << "KEY   → " << key << '\n';
@@ -54,6 +56,7 @@ static void print_value(std::string value) {
 int main() {
     using namespace kpn;
 
+    // --8<-- [start:fanout_network]
     auto gen  = make_node<generate>(out<"kv">{},          4);
     auto par  = make_node<parse>   (in<"kv">{},  out<"key", "value">{}, 4);
     auto keys = make_node<print_key>  (in<"key">{},   4);
@@ -72,4 +75,5 @@ int main() {
     net.start();
     std::this_thread::sleep_for(std::chrono::milliseconds(600));
     net.stop();
+    // --8<-- [end:fanout_network]
 }

examples/04_storage_policy/main.cpp

@@ -34,12 +34,14 @@ struct Tag {
     int value = 0;
 };
 
+// --8<-- [start:storage_policy_spec]
 // Override: store Tag by value despite being a struct
 // (it's trivially copyable and small — this just makes the policy explicit)
 template<>
 struct kpn::channel_storage_policy<Tag> {
     static constexpr bool by_value = true;
 };
+// --8<-- [end:storage_policy_spec]
 
 // ── Node functions ────────────────────────────────────────────────────────────
 

examples/05_error_handling/main.cpp

@@ -45,6 +45,7 @@ int main() {
 
     Network net;
 
+    // --8<-- [start:diagnostics_handler]
     // Custom diagnostics handler — fires on the watchdog interval.
     // Print a concise one-liner rather than the full table.
     net.set_diagnostics_handler([](const std::vector<NodeSnapshot>& nodes,
@@ -57,6 +58,7 @@ int main() {
                       << "overflows=" << c.overflows;
         std::cout << '\n';
     });
+    // --8<-- [end:diagnostics_handler]
 
     net.set_watchdog_interval(std::chrono::milliseconds(200));
 

examples/09_opencv_cellshade/example_hybrid.py

@@ -0,0 +1,101 @@
+"""
+09_opencv_cellshade/example_hybrid.py
+──────────────────────────────────────
+Hybrid cell-shading pipeline: C++ nodes handle capture, grayscale conversion,
+and edge detection; a Python/numpy function replaces the C++ quantise node;
+Python drives the display loop using cv2.
+
+Pipeline:
+                     ┌─[py_quantise]──────────────┐
+  [CaptureNode] ─────┤                             ├──[CompositeNode]──result──▶ cv2.imshow
+   out0=colour       └─[ToGrayNode]─[EdgesNode]───┘                  edges───▶ cv2.imshow
+   out1=grey
+
+For a pure-C++ version see main.cpp; for the C++ static-network version see
+12_static_cellshade/main.cpp.
+
+Press 'q' or Esc to stop.
+"""
+
+import sys
+import os
+
+# Adjust path to wherever CMake placed the .so
+BUILD_DIR = os.environ.get("KPN_BUILD_DIR",
+                           os.path.join(os.path.dirname(__file__),
+                                        "../../build/examples"))
+sys.path.insert(0, BUILD_DIR)
+
+import numpy as np
+import cv2
+import kpn_opencv as kpn
+
+
+# ── Python node: replace the C++ quantise with numpy ─────────────────────────
+# Receives and returns a BGR numpy array (H×W×3 uint8).
+
+def py_quantise(bgr: np.ndarray) -> np.ndarray:
+    levels = 4
+    step   = 256 // levels
+    q = (bgr.astype(np.int32) // step) * step + (step // 2)
+    return q.clip(0, 255).astype(np.uint8)
+
+
+# ── Build network ─────────────────────────────────────────────────────────────
+
+net = kpn.Network()
+
+net.add("src",   kpn.make_capture())          # out0=colour, out1=grey
+net.add_node("quant", py_quantise,            # Python node — numpy in/out
+             inputs=["mat"], outputs=["mat"])
+net.add("gray",  kpn.make_to_gray())          # in0=bgr → out0=gray
+net.add("edges", kpn.make_edges())            # in0=gray → out0=edge_mask
+net.add("comp",  kpn.make_composite())        # in0=edge_mask, in1=colour
+                                              # out0=result,   out1=edge_mask
+
+# src.colour → py_quantise
+net.connect("src",   0, "quant", 0)
+# src.grey   → to_gray
+net.connect("src",   1, "gray",  0)
+# gray       → edges
+net.connect("gray",  0, "edges", 0)
+# quantised colour → composite.colour  (input slot 1)
+net.connect("quant", 0, "comp",  1)
+# edge mask         → composite.edges  (input slot 0)
+net.connect("edges", 0, "comp",  0)
+
+net.build()
+net.start()
+
+# ── Display loop (drives GUI on this thread) ──────────────────────────────────
+
+cv2.namedWindow("Cell Shade (Python quant)", cv2.WINDOW_NORMAL)
+cv2.namedWindow("Edge Mask",                 cv2.WINDOW_NORMAL)
+cv2.resizeWindow("Cell Shade (Python quant)", 1280, 720)
+cv2.resizeWindow("Edge Mask",                  640, 360)
+
+try:
+    while True:
+        # Blocking reads — GIL released while waiting so C++ threads can run
+        result = net.read("comp", 0)   # composite frame  (BGR numpy array)
+        edges  = net.read("comp", 1)   # edge mask        (grayscale numpy array)
+
+        cv2.imshow("Cell Shade (Python quant)", result)
+        cv2.imshow("Edge Mask", cv2.cvtColor(edges, cv2.COLOR_GRAY2BGR))
+
+        key = cv2.waitKey(1)
+        if key in (ord('q'), 27):
+            break
+
+        # Check windows still open
+        try:
+            if cv2.getWindowProperty("Cell Shade (Python quant)",
+                                     cv2.WND_PROP_VISIBLE) < 1:
+                break
+        except cv2.error:
+            break
+
+finally:
+    net.stop()
+    cv2.destroyAllWindows()
+    del net  # let C++ destructor run before nanobind tears down

examples/09_opencv_cellshade/kpn_opencv.cpp

@@ -0,0 +1,177 @@
+#define KPN_BUILD_PYTHON
+#include <kpn/python/auto_bind.hpp>
+
+#include <nanobind/ndarray.h>
+
+#include <opencv2/core.hpp>
+#include <opencv2/imgproc.hpp>
+#include <opencv2/videoio.hpp>
+
+#include <chrono>
+#include <cmath>
+#include <iostream>
+#include <thread>
+#include <tuple>
+
+namespace nb = nanobind;
+using namespace kpn;
+using namespace kpn::python;
+
+// ── PythonConverter<cv::Mat> ──────────────────────────────────────────────────
+// Converts cv::Mat ↔ numpy array (uint8, HxW or HxWxC shape).
+//
+// to_python: clones the mat onto the heap; the numpy array owns it via a
+//   capsule deleter — no shared cv::Mat refcount dangling after the Variant dies.
+// from_python: calls numpy.ascontiguousarray, then clones into an owned cv::Mat.
+
+namespace kpn {
+
+template<> struct PythonConverter<cv::Mat> {
+    static constexpr const char* type_name = "mat";
+
+    static nb::object to_python(const cv::Mat& m) {
+        // Must be called with the GIL held (always true: called from read() or
+        // from within the gil_scoped_acquire block in PyNode::run_loop).
+        auto np = nb::module_::import_("numpy");
+        cv::Mat c = m.clone();  // ensure contiguous, independently owned
+        nb::bytes raw(reinterpret_cast<const char*>(c.data),
+                       c.total() * c.elemSize());
+        nb::object arr = np.attr("frombuffer")(raw, "uint8");
+        int H = c.rows, W = c.cols, C = c.channels();
+        arr = arr.attr("reshape")(
+            C > 1 ? nb::make_tuple(H, W, C) : nb::make_tuple(H, W));
+        return arr.attr("copy")();  // writable, lifetime-independent copy
+    }
+
+    static cv::Mat from_python(nb::object o) {
+        auto np = nb::module_::import_("numpy");
+        // Ensure contiguous uint8 layout (in-place if already compatible)
+        nb::object arr = np.attr("ascontiguousarray")(o, "uint8");
+        auto shape = nb::cast<std::vector<int>>(arr.attr("shape"));
+        if (shape.size() < 2 || shape.size() > 3)
+            throw std::runtime_error(
+                "cv::Mat from_python: expected 2D (H×W) or 3D (H×W×C) uint8 array");
+        int H = shape[0], W = shape[1];
+        int C = (shape.size() == 3) ? shape[2] : 1;
+        int type = C > 1 ? CV_8UC(C) : CV_8UC1;
+
+        // Cast to ndarray to get the raw data pointer
+        auto binfo = nb::cast<nb::ndarray<nb::numpy, uint8_t>>(arr);
+        cv::Mat wrap(H, W, type, binfo.data());
+        return wrap.clone();  // own the pixel data
+    }
+};
+
+} // namespace kpn
+
+// ── Pipeline functions ────────────────────────────────────────────────────────
+
+static cv::Mat make_gradient(int W, int H) {
+    cv::Mat xr(H, W, CV_8UC1), yg(H, W, CV_8UC1), b(H, W, CV_8UC1, cv::Scalar(128));
+    for (int x = 0; x < W; ++x) xr.col(x).setTo(x * 255 / W);
+    for (int y = 0; y < H; ++y) yg.row(y).setTo(y * 255 / H);
+    cv::Mat channels[3] = {b, yg, xr};
+    cv::Mat grad;
+    cv::merge(channels, 3, grad);
+    return grad;
+}
+
+static std::tuple<cv::Mat, cv::Mat> capture() {
+    constexpr int W = 640, H = 480;
+    static cv::VideoCapture cap;
+    static bool opened = false;
+    if (!opened) {
+        opened = true;
+        cap.open(0, cv::CAP_V4L2);
+        if (cap.isOpened()) {
+            cap.set(cv::CAP_PROP_FRAME_WIDTH,  W);
+            cap.set(cv::CAP_PROP_FRAME_HEIGHT, H);
+        } else {
+            std::cerr << "[capture] no webcam — using synthetic animated pattern\n";
+        }
+    }
+
+    cv::Mat frame;
+    if (cap.isOpened()) {
+        auto t0 = std::chrono::steady_clock::now();
+        cap >> frame;
+        auto elapsed = std::chrono::steady_clock::now() - t0;
+        if (elapsed < std::chrono::milliseconds(20))
+            std::this_thread::sleep_for(std::chrono::milliseconds(33) - elapsed);
+        if (frame.empty()) frame = cv::Mat::zeros(H, W, CV_8UC3);
+    } else {
+        static int tick = 0;
+        static cv::Mat grad = make_gradient(W, H);
+        ++tick;
+        frame = grad.clone();
+        int r = 150 + (tick % 80) * 4;
+        cv::circle(frame, {W/2,   H/2},   r,     {255, 200,   0}, -1);
+        cv::circle(frame, {W/2,   H/2},   r / 2, {  0, 128, 255}, -1);
+        cv::circle(frame, {W*2/5, H*2/5}, r / 3, {200,   0, 200}, -1);
+        std::this_thread::sleep_for(std::chrono::milliseconds(33));
+    }
+    return {frame.clone(), frame.clone()};
+}
+
+static cv::Mat to_gray(cv::Mat bgr) {
+    cv::Mat gray;
+    cv::cvtColor(bgr, gray, cv::COLOR_BGR2GRAY);
+    return gray;
+}
+
+static cv::Mat edges_fn(cv::Mat gray) {
+    cv::Mat blurred, mask;
+    cv::GaussianBlur(gray, blurred, {5, 5}, 0);
+    cv::Canny(blurred, mask, 50, 150);
+    return mask;
+}
+
+static cv::Mat quantise(cv::Mat bgr) {
+    constexpr int levels = 4;
+    constexpr double step = 256.0 / levels;
+    static const cv::Mat lut = []() {
+        cv::Mat l(1, 256, CV_8UC1);
+        for (int i = 0; i < 256; ++i)
+            l.at<uchar>(i) = cv::saturate_cast<uchar>(
+                std::floor(i / step) * step + step / 2.0);
+        return l;
+    }();
+    cv::Mat out;
+    cv::LUT(bgr, lut, out);
+    return out;
+}
+
+// Returns composite frame AND edge mask so the display node can show both
+// without needing a fan-out on the edges channel.
+static std::tuple<cv::Mat, cv::Mat> composite(cv::Mat edge_mask, cv::Mat colour) {
+    cv::Mat result = colour.clone();
+    result.setTo(cv::Scalar(0, 0, 0), edge_mask);
+    return {result, edge_mask};
+}
+
+// ── Registry ──────────────────────────────────────────────────────────────────
+// Variant deduced as std::variant<cv::Mat> — every node uses only cv::Mat.
+
+using CvNodes = NodeRegistry<
+    Entry<capture,   "capture">,
+    Entry<to_gray,   "to_gray">,
+    Entry<edges_fn,  "edges">,
+    Entry<quantise,  "quantise">,
+    Entry<composite, "composite">
+>;
+
+// ── Module ────────────────────────────────────────────────────────────────────
+
+NB_MODULE(kpn_opencv, m) {
+    m.doc() = "KPN++ OpenCV bindings for the cell-shading pipeline";
+
+    // Registers: Network, INode, CaptureNode, ToGrayNode, EdgesNode,
+    //            QuantiseNode, CompositeNode, and make_<name>() factories.
+    // Network.add_node(name, callable, inputs=["mat"], outputs=["mat"])
+    // accepts Python callables that receive/return numpy uint8 arrays.
+    bind_network<CvNodes>(m);
+
+    // Note: bind_debug is omitted here — cv::Mat functions cannot be called
+    // directly from Python without the variant/network machinery. Use
+    // net.write() + net.read() to inject/inspect individual nodes instead.
+}

examples/09_opencv_cellshade/main.cpp

@@ -8,6 +8,12 @@
 #include <thread>
 #include <chrono>
 
+// Teach KPN how many bytes a cv::Mat actually carries (header + pixel data).
+template<>
+struct kpn::ChannelDataSize<cv::Mat> {
+    static std::size_t bytes(const cv::Mat& m) { return m.total() * m.elemSize(); }
+};
+
 // ── Cell-shading pipeline ─────────────────────────────────────────────────────
 //
 //  [capture] --"colour"--> [quantise] ──────────────────────────┐
@@ -32,6 +38,7 @@ static cv::Mat make_gradient(int W, int H) {
 
 // ── Pipeline functions ────────────────────────────────────────────────────────
 
+// --8<-- [start:capture_fn]
 static std::tuple<cv::Mat, cv::Mat> capture() {
     constexpr int W = 640, H = 480;
     static cv::VideoCapture cap;
@@ -68,6 +75,7 @@ static std::tuple<cv::Mat, cv::Mat> capture() {
     }
     return {frame.clone(), frame.clone()};
 }
+// --8<-- [end:capture_fn]
 
 static cv::Mat to_gray(cv::Mat bgr) {
     cv::Mat gray;
@@ -112,6 +120,7 @@ static std::tuple<cv::Mat, cv::Mat> composite(cv::Mat edge_mask, cv::Mat colour)
 // The constructor opens both windows on the main thread (Wayland requirement).
 // operator() is called by step() whenever both channels have a frame ready.
 
+// --8<-- [start:display_node]
 class DisplayNode : public kpn::MainThreadNode<DisplayNode,
                                                kpn::in<"composite", "edges">,
                                                cv::Mat, cv::Mat> {
@@ -141,12 +150,14 @@ private:
         catch (const cv::Exception&) { return false; }
     }
 };
+// --8<-- [end:display_node]
 
 // ─────────────────────────────────────────────────────────────────────────────
 
 int main() {
     using namespace kpn;
 
+    // --8<-- [start:opencv_network]
     auto src       = make_node<capture>  (out<"colour","grey">{},            8);
     auto gray_node = make_node<to_gray>  (in<"bgr">{},   out<"gray">{},      8);
     auto edge_node = make_node<edges_fn> (in<"gray">{},  out<"edges">{},     8);
@@ -171,13 +182,17 @@ int main() {
        .connect("comp",  comp.template output<"result">(),     "display", disp.template input<"composite">())
        .connect("comp",  comp.template output<"edges">(),      "display", disp.template input<"edges">())
        .build();
+    // --8<-- [end:opencv_network]
 
     net.set_watchdog_interval(std::chrono::milliseconds(5000));
+#ifdef KPN_WEB_DEBUG
     net.set_web_debug_port(9090);
+#endif
 
     std::cout << "Cell-shading pipeline running. Press 'q' to stop.\n";
     std::cout << "Web debug UI: http://localhost:9090\n";
 
+    // --8<-- [start:main_thread_step]
     net.start();
 
     // Main thread drives display — imshow/waitKey stay on the GUI thread.
@@ -186,5 +201,6 @@ int main() {
         cv::waitKey(8); // yield event loop when no frame ready
 
     net.stop();
+    // --8<-- [end:main_thread_step]
     return 0;
 }

examples/14_debug_hub/main.cpp

@@ -0,0 +1,159 @@
+// Example 14 — DebugHub with Shared Resource Token
+//
+// Two independent KPN networks compete for one shared inference resource
+// (simulating a small GPU or single-session ONNX runtime).  The DebugHub
+// serves a single web UI at http://localhost:9090 with:
+//
+//   [All Networks]  — resource utilisation cards + cross-network node table
+//   [detect]        — force-directed graph for the detection pipeline
+//   [classify]      — force-directed graph for the classification pipeline
+//
+// Topology:
+//
+//   detect pipeline:
+//     [source_detect] ──> [run_detect] ──> [sink_detect]
+//
+//   classify pipeline:
+//     [source_classify] ──> [run_classify] ──> [sink_classify]
+//
+// Both [run_detect] and [run_classify] call gpu.acquire() before touching the
+// simulated device.  The priority-based token awards the next slot to the
+// waiter that is more likely to make useful progress (higher priority score).
+//
+// Build:  cmake -DKPN_WEB_DEBUG=ON ..  &&  cmake --build .
+// Run:    ./14_debug_hub
+// UI:     http://localhost:9090
+
+#ifdef KPN_WEB_DEBUG
+
+#include <kpn/kpn.hpp>
+
+#include <atomic>
+#include <chrono>
+#include <iostream>
+#include <thread>
+
+using namespace kpn;
+using namespace std::chrono_literals;
+
+// ── Simulated inference device ────────────────────────────────────────────────
+//
+// Represents any exclusive, serialised accelerator: GPU session, ONNX runtime,
+// hardware encoder, etc.  Only one caller can hold it at a time.
+
+struct GPU {
+    // Detection model: fast, 8 ms per frame.
+    int detect(int frame_id) {
+        std::this_thread::sleep_for(8ms);
+        return frame_id * 2;           // synthetic "score"
+    }
+
+    // Classification model: heavier, 14 ms per frame.
+    int classify(int frame_id) {
+        std::this_thread::sleep_for(14ms);
+        return frame_id % 10;          // synthetic "label"
+    }
+};
+
+// Global pointer so free-function nodes can reach the resource.
+// In production code, capture by reference inside an ObjectNode functor instead.
+static SharedResource<GPU>* g_gpu = nullptr;
+
+// ── Detection pipeline ────────────────────────────────────────────────────────
+
+static int source_detect() {
+    static std::atomic<int> id{0};
+    std::this_thread::sleep_for(25ms);      // ~40 fps source rate
+    return id.fetch_add(1, std::memory_order_relaxed);
+}
+
+static int run_detect(int frame_id) {
+    // Higher priority: detection is latency-critical.
+    auto guard = g_gpu->acquire([] { return 0.7f; });
+    return guard->detect(frame_id);
+}
+
+static std::atomic<uint64_t> detect_out{0};
+static void sink_detect(int) {
+    detect_out.fetch_add(1, std::memory_order_relaxed);
+}
+
+// ── Classification pipeline ───────────────────────────────────────────────────
+
+static int source_classify() {
+    static std::atomic<int> id{0};
+    std::this_thread::sleep_for(40ms);      // ~25 fps source rate
+    return id.fetch_add(1, std::memory_order_relaxed);
+}
+
+static int run_classify(int frame_id) {
+    // Lower priority: classification is best-effort.
+    auto guard = g_gpu->acquire([] { return 0.3f; });
+    return guard->classify(frame_id);
+}
+
+static std::atomic<uint64_t> classify_out{0};
+static void sink_classify(int) {
+    classify_out.fetch_add(1, std::memory_order_relaxed);
+}
+
+// ── main ──────────────────────────────────────────────────────────────────────
+
+int main() {
+    SharedResource<GPU> gpu;
+    g_gpu = &gpu;
+
+    // ── Detection network ─────────────────────────────────────────────────────
+    auto src_det = make_node<source_detect, "source_detect">(4);
+    auto inf_det = make_node<run_detect,    "run_detect"   >(4);
+    auto snk_det = make_node<sink_detect,   "sink_detect"  >(4);
+
+    auto net_detect = make_network(
+        edge(src_det.output<0>(), inf_det.input<0>()),
+        edge(inf_det.output<0>(), snk_det.input<0>())
+    );
+
+    // ── Classification network ────────────────────────────────────────────────
+    auto src_cls = make_node<source_classify, "source_classify">(4);
+    auto inf_cls = make_node<run_classify,    "run_classify"   >(4);
+    auto snk_cls = make_node<sink_classify,   "sink_classify"  >(4);
+
+    auto net_classify = make_network(
+        edge(src_cls.output<0>(), inf_cls.input<0>()),
+        edge(inf_cls.output<0>(), snk_cls.input<0>())
+    );
+
+    // ── Hub — one debug server for both networks + the shared resource ─────────
+    DebugHub hub(9090);
+    hub.register_network("detect",   net_detect);
+    hub.register_network("classify", net_classify);
+    hub.register_resource("gpu",     &gpu);
+
+    net_detect.start();
+    net_classify.start();
+    hub.start();
+
+    std::cout << "Running — open http://localhost:9090\n"
+              << "Tabs: [All Networks]  [detect]  [classify]\n"
+              << "Press Enter to stop.\n";
+    std::cin.get();
+
+    net_detect.stop();
+    net_classify.stop();
+
+    std::cout << "\nResults:\n"
+              << "  detect:   " << detect_out.load()   << " frames\n"
+              << "  classify: " << classify_out.load() << " frames\n";
+    return 0;
+}
+
+#else  // no KPN_WEB_DEBUG
+
+#include <iostream>
+int main() {
+    std::cerr << "This example requires KPN_WEB_DEBUG.\n"
+              << "Rebuild with:  cmake -DKPN_WEB_DEBUG=ON ..\n";
+    return 1;
+}
+
+#endif

examples/15_node_error_handler/main.cpp

@@ -0,0 +1,75 @@
+// Example 15 — Per-node Error Handler
+//
+// Demonstrates set_error_handler() for deciding whether a network can
+// continue when a node throws an exception.
+//
+// The "validator" node rejects even numbers by throwing std::runtime_error.
+// Its error handler logs the failure and returns true (skip & continue),
+// so odd numbers still flow through to the sink.
+//
+// Compare: a second handler (commented below) returns false instead,
+// which stops the node and gracefully shuts the downstream side down.
+//
+// Pipeline:  [source] --int--> [validator] --int--> [sink]
+
+#include <kpn/kpn.hpp>
+#include <chrono>
+#include <iostream>
+#include <thread>
+
+static int counter = 0;
+
+static int source() {
+    std::this_thread::sleep_for(std::chrono::milliseconds(20));
+    return ++counter;
+}
+
+static int validate(int x) {
+    if (x % 2 == 0)
+        throw std::runtime_error("even number rejected: " + std::to_string(x));
+    return x;
+}
+
+static int received = 0;
+
+static void sink(int x) {
+    std::cout << "  processed: " << x << '\n';
+    ++received;
+}
+
+int main() {
+    using namespace kpn;
+
+    auto src  = make_node<source>  ();
+    auto proc = make_node<validate>();
+    auto snk  = make_node<sink>   ();
+
+    // --8<-- [start:error_handler]
+    // Return true  → skip this invocation, keep the node running.
+    // Return false → stop the node (downstream drains then also stops).
+    proc.set_error_handler([](std::string_view name, std::exception_ptr ep) {
+        try { std::rethrow_exception(ep); }
+        catch (const std::exception& e) {
+            std::cerr << "[" << name << "] skipping item — " << e.what() << '\n';
+        }
+        return true;
+    });
+    // --8<-- [end:error_handler]
+
+    Network net;
+    net.add("source",    src)
+       .add("validator", proc)
+       .add("sink",      snk)
+       .connect("source",    src.output<0>(),  "validator", proc.input<0>())
+       .connect("validator", proc.output<0>(), "sink",      snk.input<0>())
+       .build();
+
+    std::cout << "source emits 1..N; validator rejects even numbers.\n"
+              << "Error messages on stderr, accepted items on stdout.\n\n";
+
+    net.start();
+    std::this_thread::sleep_for(std::chrono::milliseconds(300));
+    net.stop();
+
+    std::cout << "\nItems accepted by sink: " << received << '\n';
+}

examples/16_event_callbacks/main.cpp

@@ -0,0 +1,83 @@
+// Example 16 — Event Callbacks: overflow and node-stopped signals
+//
+// Two complementary observation mechanisms:
+//
+// 1. Per-node overflow callback   set_overflow_callback()
+//    Fired (with a timestamp) when a node's output channel is full and an
+//    item is dropped.  Useful for targeted monitoring of a specific node.
+//
+// 2. Network-level event handler  net.set_event_handler()
+//    Aggregate callback covering every node: receives the node name, a
+//    NodeEvent (Overflow or Closed), and a timestamp.  Register once and
+//    observe the whole network.
+//
+// Pipeline:  [fast_source] --int--> [slow_sink]
+//
+// fast_source produces at ~500 items/s; slow_sink consumes at ~20 items/s.
+// The channel capacity is 3, so overflows appear within milliseconds.
+
+#include <kpn/kpn.hpp>
+#include <atomic>
+#include <chrono>
+#include <iostream>
+#include <thread>
+
+using namespace kpn;
+using namespace std::chrono;
+
+// ── Node functions ────────────────────────────────────────────────────────────
+
+// --8<-- [start:node_fns]
+static std::atomic<int> g_seq{0};
+
+static int fast_source() {
+    std::this_thread::sleep_for(milliseconds(2));   // ~500/s
+    return g_seq.fetch_add(1);
+}
+
+static void slow_sink(int x) {
+    std::this_thread::sleep_for(milliseconds(50));  // ~20/s
+    std::cout << "  consumed: " << x << '\n';
+}
+// --8<-- [end:node_fns]
+
+// ── main ──────────────────────────────────────────────────────────────────────
+
+int main() {
+    auto src = make_node<fast_source>(/*capacity=*/3);
+    auto snk = make_node<slow_sink> (/*capacity=*/3);
+
+    // --8<-- [start:per_node_callback]
+    // Per-node overflow callback — no node name needed, known at registration.
+    std::atomic<int> overflow_count{0};
+    src.set_overflow_callback([&](steady_clock::time_point ts) {
+        auto ms = duration_cast<milliseconds>(ts.time_since_epoch()).count();
+        std::cerr << "[overflow] fast_source at t=" << ms << "ms\n";
+        overflow_count.fetch_add(1);
+    });
+    // --8<-- [end:per_node_callback]
+
+    Network net;
+
+    // --8<-- [start:network_event_handler]
+    // Network-level aggregate handler — covers every node, includes node name.
+    net.set_event_handler([](std::string_view name, NodeEvent ev,
+                             steady_clock::time_point ts) {
+        auto ms = duration_cast<milliseconds>(ts.time_since_epoch()).count();
+        std::string_view kind = (ev == NodeEvent::Overflow) ? "overflow" : "closed";
+        std::cerr << "[net:" << kind << "] node=" << name << " t=" << ms << "ms\n";
+    });
+    // --8<-- [end:network_event_handler]
+
+    net.add("source", src)
+       .add("sink",   snk)
+       .connect("source", src.output<0>(), "sink", snk.input<0>())
+       .build()
+       .start();
+
+    std::this_thread::sleep_for(milliseconds(300));
+    net.stop();
+
+    std::cout << "\nTotal overflows observed by per-node callback: "
+              << overflow_count.load() << '\n';
+}

examples/CMakeLists.txt

@@ -1,8 +1,16 @@
 cmake_minimum_required(VERSION 3.21)
 
+# Build an example and register it as a CTest smoke test.
+# Examples that are self-terminating (fixed sleep → net.stop()) pass when
+# they exit 0 within TIMEOUT seconds. OpenCV/UI examples are excluded.
 function(kpn_example name)
     add_executable(${name} ${name}/main.cpp)
     target_link_libraries(${name} PRIVATE kpn)
+    add_test(NAME example_${name} COMMAND ${name})
+    set_tests_properties(example_${name} PROPERTIES
+        TIMEOUT 15
+        LABELS  examples
+    )
 endfunction()
 
 kpn_example(01_hello_pipeline)
@@ -11,19 +19,30 @@ kpn_example(03_multi_output)
 kpn_example(04_storage_policy)
 kpn_example(05_error_handling)
 kpn_example(06_watchdog)
+set_tests_properties(example_06_watchdog PROPERTIES TIMEOUT 40)
 kpn_example(10_static_hello_pipeline)
 kpn_example(11_static_fanout)
+kpn_example(15_node_error_handler)
+kpn_example(16_event_callbacks)
 if(KPN_WEB_DEBUG)
     kpn_target_enable_web_debug(06_watchdog)
+
+    add_executable(14_debug_hub 14_debug_hub/main.cpp)
+    target_link_libraries(14_debug_hub PRIVATE kpn)
+    kpn_target_enable_web_debug(14_debug_hub)
 endif()
-# 07 and 08 require the Python bindings — only add if built
-if(KPN_BUILD_PYTHON)
-    # These are Python scripts, not compiled targets — installed alongside kpn_python
-endif()
+# 07 and 08 are Python scripts — no compiled target needed.
 
 # 09 requires OpenCV — only build if found
 find_package(OpenCV QUIET COMPONENTS core imgproc highgui videoio)
 if(OpenCV_FOUND)
+    # Hybrid Python example: kpn_opencv module (requires both OpenCV and nanobind)
+    if(KPN_BUILD_PYTHON)
+        nanobind_add_module(kpn_opencv 09_opencv_cellshade/kpn_opencv.cpp)
+        target_link_libraries(kpn_opencv PRIVATE kpn ${OpenCV_LIBS})
+        target_compile_definitions(kpn_opencv PRIVATE KPN_BUILD_PYTHON)
+        message(STATUS "KPN++ kpn_opencv Python module: building (OpenCV ${OpenCV_VERSION})")
+    endif()
     add_executable(09_opencv_cellshade 09_opencv_cellshade/main.cpp)
     target_link_libraries(09_opencv_cellshade PRIVATE kpn ${OpenCV_LIBS})
 

include/kpn/branch.hpp

@@ -0,0 +1,301 @@
+#pragma once
+#include "channel.hpp"
+#include "diagnostics.hpp"
+#include "inode.hpp"
+#include "port.hpp"
+#include "traits.hpp"
+
+#include <array>
+#include <atomic>
+#include <functional>
+#include <memory>
+#include <thread>
+
+namespace kpn {
+
+// ── RouterNode ────────────────────────────────────────────────────────────────
+//
+// Reads one item and pushes it to exactly one of N output channels, chosen by
+// selector(item). If selector returns >= N the item is silently dropped.
+//
+// Usage:
+//   auto router = make_router<Image, 3>(
+//       [](const Image& img) -> std::size_t { return img.stream_id % 3; });
+//   net.connect("src",    src.output<0>(),     "router", router.input<0>())
+//      .connect("router", router.output<0>(),  "nodeA",  nodeA.input<0>())
+//      .connect("router", router.output<1>(),  "nodeB",  nodeB.input<0>())
+//      .connect("router", router.output<2>(),  "nodeC",  nodeC.input<0>());
+
+template<typename T, std::size_t N, std::size_t Id = 0>
+class RouterNode : public INode {
+public:
+    using Selector     = std::function<std::size_t(const T&)>;
+    using args_tuple   = std::tuple<T>;
+    using return_tuple = repeat_tuple_t<T, N>;
+    using return_raw   = return_tuple;
+
+    static constexpr std::size_t input_count    = 1;
+    static constexpr std::size_t output_count   = N;
+    static constexpr std::size_t unique_tag     = Id;
+    static constexpr bool        is_router_node = true;
+
+    explicit RouterNode(Selector sel, std::size_t fifo_capacity = 5)
+        : selector_(std::move(sel))
+        , fifo_capacity_(fifo_capacity)
+    {
+        input_ch_ = std::make_shared<Channel<T>>(fifo_capacity);
+    }
+
+    ~RouterNode() override { stop(); }
+
+    // ── INode ─────────────────────────────────────────────────────────────────
+
+    void start() override {
+        input_ch_->enable();
+        stop_flag_.store(false, std::memory_order_relaxed);
+        thread_ = std::jthread([this](std::stop_token) { run_loop(); });
+    }
+
+    void stop() override {
+        stop_flag_.store(true, std::memory_order_relaxed);
+        input_ch_->disable();
+        if (thread_.joinable()) thread_.request_stop(), thread_.join();
+    }
+
+    bool running() const override {
+        return thread_.joinable() && !stop_flag_.load(std::memory_order_relaxed);
+    }
+
+    void set_name(std::string name) override { name_ = std::move(name); }
+
+    const NodeStats& stats() const override { return stats_; }
+
+    NodeSnapshot node_snapshot(const std::string& name, double elapsed_s) const override {
+        uint64_t frames   = stats_.frames_processed.load(std::memory_order_relaxed);
+        double exec_ms    = stats_.ema_exec_us.load(std::memory_order_relaxed) / 1000.0;
+        double blocked_ms = stats_.total_blocked_us.load(std::memory_order_relaxed) / 1000.0;
+        double total_ms   = exec_ms + blocked_ms;
+        return {name, frames, exec_ms,
+                stats_.max_exec_us.load(std::memory_order_relaxed) / 1000.0,
+                blocked_ms,
+                elapsed_s > 0 ? frames / elapsed_s : 0.0,
+                stats_.total_cpu_us.load(std::memory_order_relaxed) / 1000.0,
+                total_ms > 0 ? 100.0 * exec_ms / total_ms : 0.0};
+    }
+
+    // ── Port access ───────────────────────────────────────────────────────────
+
+    template<std::size_t I = 0>
+    InputPort<RouterNode, I> input() {
+        static_assert(I == 0, "RouterNode has exactly one input");
+        return {*this};
+    }
+
+    template<std::size_t I>
+    OutputPort<RouterNode, I> output() {
+        static_assert(I < N, "RouterNode output index out of range");
+        return {*this};
+    }
+
+    // ── Internal channel accessors (called by Network::connect) ───────────────
+
+    template<std::size_t I>
+    Channel<T>& input_channel() {
+        static_assert(I == 0);
+        return *input_ch_;
+    }
+
+    template<std::size_t I>
+    void set_input_channel(std::shared_ptr<Channel<T>> ch) {
+        static_assert(I == 0);
+        input_ch_ = std::move(ch);
+    }
+
+    template<std::size_t I>
+    void set_output_channel(Channel<T>* ch) {
+        static_assert(I < N);
+        out_channels_[I] = ch;
+    }
+
+private:
+    void run_loop() {
+        while (!stop_flag_.load(std::memory_order_relaxed)) {
+            try {
+                auto        t0  = clock_t::now();
+                T           val = input_ch_->pop();
+                auto        t1  = clock_t::now();
+                auto        cpu0 = NodeStats::cpu_now();
+
+                std::size_t idx = selector_(val);
+                if (idx < N && out_channels_[idx]) {
+                    try { out_channels_[idx]->push(val); }
+                    catch (const ChannelOverflowError&) {}
+                }
+
+                auto cpu1 = NodeStats::cpu_now();
+                auto t2   = clock_t::now();
+                stats_.record_exec(duration_t(t2 - t1), duration_t(t1 - t0), cpu0, cpu1);
+            } catch (const ChannelClosedError&) {
+                break;
+            }
+        }
+    }
+
+    std::string                 name_;
+    std::size_t                 fifo_capacity_;
+    Selector                    selector_;
+    std::shared_ptr<Channel<T>> input_ch_;
+    std::array<Channel<T>*, N>  out_channels_{};
+    std::atomic<bool>           stop_flag_{false};
+    std::jthread                thread_;
+    NodeStats                   stats_;
+};
+
+// ── FilterNode ────────────────────────────────────────────────────────────────
+//
+// Reads one item and pushes it downstream only when pred(item) is true.
+// Dropped items are not counted as processed frames.
+//
+// Usage:
+//   auto filt = make_filter<Frame>([](const Frame& f) { return f.valid; });
+//   net.connect("src",  src.output<0>(),  "filt", filt.input<0>())
+//      .connect("filt", filt.output<0>(), "dst",  dst.input<0>());
+
+template<typename T, std::size_t Id = 0>
+class FilterNode : public INode {
+public:
+    using Predicate    = std::function<bool(const T&)>;
+    using args_tuple   = std::tuple<T>;
+    using return_tuple = std::tuple<T>;
+    using return_raw   = return_tuple;
+
+    static constexpr std::size_t input_count    = 1;
+    static constexpr std::size_t output_count   = 1;
+    static constexpr std::size_t unique_tag     = Id;
+    static constexpr bool        is_filter_node = true;
+
+    explicit FilterNode(Predicate pred, std::size_t fifo_capacity = 5)
+        : pred_(std::move(pred))
+        , fifo_capacity_(fifo_capacity)
+    {
+        input_ch_ = std::make_shared<Channel<T>>(fifo_capacity);
+    }
+
+    ~FilterNode() override { stop(); }
+
+    // ── INode ─────────────────────────────────────────────────────────────────
+
+    void start() override {
+        input_ch_->enable();
+        stop_flag_.store(false, std::memory_order_relaxed);
+        thread_ = std::jthread([this](std::stop_token) { run_loop(); });
+    }
+
+    void stop() override {
+        stop_flag_.store(true, std::memory_order_relaxed);
+        input_ch_->disable();
+        if (thread_.joinable()) thread_.request_stop(), thread_.join();
+    }
+
+    bool running() const override {
+        return thread_.joinable() && !stop_flag_.load(std::memory_order_relaxed);
+    }
+
+    void set_name(std::string name) override { name_ = std::move(name); }
+
+    const NodeStats& stats() const override { return stats_; }
+
+    NodeSnapshot node_snapshot(const std::string& name, double elapsed_s) const override {
+        uint64_t frames   = stats_.frames_processed.load(std::memory_order_relaxed);
+        double exec_ms    = stats_.ema_exec_us.load(std::memory_order_relaxed) / 1000.0;
+        double blocked_ms = stats_.total_blocked_us.load(std::memory_order_relaxed) / 1000.0;
+        double total_ms   = exec_ms + blocked_ms;
+        return {name, frames, exec_ms,
+                stats_.max_exec_us.load(std::memory_order_relaxed) / 1000.0,
+                blocked_ms,
+                elapsed_s > 0 ? frames / elapsed_s : 0.0,
+                stats_.total_cpu_us.load(std::memory_order_relaxed) / 1000.0,
+                total_ms > 0 ? 100.0 * exec_ms / total_ms : 0.0};
+    }
+
+    // ── Port access ───────────────────────────────────────────────────────────
+
+    template<std::size_t I = 0>
+    InputPort<FilterNode, I> input() {
+        static_assert(I == 0, "FilterNode has exactly one input");
+        return {*this};
+    }
+
+    template<std::size_t I = 0>
+    OutputPort<FilterNode, I> output() {
+        static_assert(I == 0, "FilterNode has exactly one output");
+        return {*this};
+    }
+
+    // ── Internal channel accessors (called by Network::connect) ───────────────
+
+    template<std::size_t I>
+    Channel<T>& input_channel() {
+        static_assert(I == 0);
+        return *input_ch_;
+    }
+
+    template<std::size_t I>
+    void set_input_channel(std::shared_ptr<Channel<T>> ch) {
+        static_assert(I == 0);
+        input_ch_ = std::move(ch);
+    }
+
+    template<std::size_t I>
+    void set_output_channel(Channel<T>* ch) {
+        static_assert(I == 0);
+        out_ch_ = ch;
+    }
+
+private:
+    void run_loop() {
+        while (!stop_flag_.load(std::memory_order_relaxed)) {
+            try {
+                auto t0  = clock_t::now();
+                T    val = input_ch_->pop();
+                auto t1  = clock_t::now();
+                auto cpu0 = NodeStats::cpu_now();
+
+                if (pred_(val) && out_ch_) {
+                    try { out_ch_->push(val); }
+                    catch (const ChannelOverflowError&) {}
+                    auto cpu1 = NodeStats::cpu_now();
+                    auto t2   = clock_t::now();
+                    stats_.record_exec(duration_t(t2 - t1), duration_t(t1 - t0), cpu0, cpu1);
+                }
+            } catch (const ChannelClosedError&) {
+                break;
+            }
+        }
+    }
+
+    std::string                 name_;
+    std::size_t                 fifo_capacity_;
+    Predicate                   pred_;
+    std::shared_ptr<Channel<T>> input_ch_;
+    Channel<T>*                 out_ch_{nullptr};
+    std::atomic<bool>           stop_flag_{false};
+    std::jthread                thread_;
+    NodeStats                   stats_;
+};
+
+// ── Factories ─────────────────────────────────────────────────────────────────
+
+template<typename T, std::size_t N>
+RouterNode<T, N> make_router(std::function<std::size_t(const T&)> sel,
+                              std::size_t capacity = 5) {
+    return RouterNode<T, N, 0>(std::move(sel), capacity);
+}
+
+template<typename T>
+FilterNode<T> make_filter(std::function<bool(const T&)> pred,
+                           std::size_t capacity = 5) {
+    return FilterNode<T, 0>(std::move(pred), capacity);
+}
+
+} // namespace kpn

include/kpn/channel.hpp

@@ -2,16 +2,30 @@
 #include "diagnostics.hpp"
 #include <atomic>
 #include <chrono>
-#include <condition_variable>
+#include <cstdint>
+#include <functional>
 #include <memory>
-#include <mutex>
-#include <queue>
 #include <stdexcept>
 #include <string>
+#include <thread>
 #include <type_traits>
 
 namespace kpn {
 
+// ── Data size trait ───────────────────────────────────────────────────────────
+// Returns the number of bytes of logical payload carried by a value.
+// Defaults to sizeof(T), which is correct for PODs and fixed-size types.
+// Specialize for heap-owning types (e.g. cv::Mat) to get accurate bandwidth:
+//
+//   template<> struct kpn::ChannelDataSize<cv::Mat> {
+//       static std::size_t bytes(const cv::Mat& m) { return m.total() * m.elemSize(); }
+//   };
+
+template<typename T>
+struct ChannelDataSize {
+    static std::size_t bytes(const T&) { return sizeof(T); }
+};
+
 // ── Storage policy ────────────────────────────────────────────────────────────
 
 template<typename T>
@@ -44,67 +58,143 @@ public:
     ChannelClosedError() : std::runtime_error("channel closed") {}
 };
 
+// ── CPU pause hint ────────────────────────────────────────────────────────────
+// Signals the CPU that this is a spin-wait loop, improving HT sibling throughput
+// and preventing branch-predictor thrash on x86. Falls back to a compiler barrier.
+
+[[maybe_unused]] static void spin_hint() noexcept {
+#if defined(__x86_64__) || defined(__i386__)
+    __asm__ volatile("pause" ::: "memory");
+#elif defined(__aarch64__) || defined(__arm__)
+    __asm__ volatile("yield" ::: "memory");
+#else
+    std::atomic_signal_fence(std::memory_order_seq_cst);
+#endif
+}
+
 // ── Channel ───────────────────────────────────────────────────────────────────
+// SPSC ring buffer with atomic wait/notify and configurable spin-before-sleep.
+//
+// `spin_count` (constructor arg, default 200): number of pause-hint iterations
+// before falling back to atomic::wait (futex). At ~20 ns/pause on x86 this is
+// ~4 µs. Set to 0 to disable spinning (useful for power-constrained or
+// predominantly-idle pipelines).
+//
+// Memory ordering contract (SPSC):
+//   push(): tail_.store(release) pairs with pop()'s tail_.load(acquire)
+//           head_.load(acquire)  pairs with pop()'s head_.store(release)
+//   pop():  head_.store(release) pairs with push()'s head_.load(acquire)
+//           tail_.load(acquire)  pairs with push()'s tail_.store(release)
 
 template<typename T>
 class Channel {
 public:
     using storage_type = channel_storage_t<T>;
 
-    explicit Channel(std::size_t capacity = 5) : capacity_(capacity) {}
+    explicit Channel(std::size_t capacity = 5, std::size_t spin_count = 200)
+        : capacity_(capacity), spin_count_(spin_count)
+    {
+        std::size_t rs = 1;
+        while (rs <= capacity) rs <<= 1;   // smallest power-of-2 > capacity
+        ring_mask_ = rs - 1;
+        buf_ = std::make_unique<storage_type[]>(rs);
+    }
 
     Channel(const Channel&)            = delete;
     Channel& operator=(const Channel&) = delete;
 
     // Push a value.
-    // - If channel is disabled (accepting_ == false): silently drop, return immediately.
-    // - If channel is full: throw ChannelOverflowError.
+    // - If channel is disabled (accepting_ == false): silently drop.
+    // - If channel is full (fill >= capacity_): throw ChannelOverflowError.
     void push(T value) {
         if (!accepting_.load(std::memory_order_relaxed)) {
             stats_.record_drop();
             return;
         }
-        std::unique_lock lock(mutex_);
-        if (!accepting_.load(std::memory_order_relaxed)) {
+        const std::size_t data_bytes = ChannelDataSize<T>::bytes(value);
+        const std::size_t t = tail_.load(std::memory_order_relaxed);
+        const std::size_t h = head_.load(std::memory_order_acquire);
+
+        if (!accepting_.load(std::memory_order_acquire)) {
             stats_.record_drop();
             return;
         }
-        if (queue_.size() >= capacity_) {
+        if (t - h >= capacity_) {
             stats_.record_overflow();
             throw ChannelOverflowError(capacity_);
         }
-        queue_.push(make_storage(std::move(value)));
-        stats_.record_push(queue_.size());
-        lock.unlock();
-        cv_.notify_one();
+
+        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_();
     }
 
-    // Blocking pop. Unblocks when an item is available or the channel is disabled.
-    // Throws ChannelClosedError if disabled and queue is empty.
+    // Blocking pop. Returns when an item is available.
+    // Throws ChannelClosedError if the channel is disabled (regardless of fill).
     T pop() {
-        std::unique_lock lock(mutex_);
-        cv_.wait(lock, [this] {
-            return !queue_.empty() || !accepting_.load(std::memory_order_relaxed);
-        });
-        if (queue_.empty())
-            throw ChannelClosedError{};
-        T value = extract(std::move(queue_.front()));
-        queue_.pop();
-        stats_.record_pop();
-        return value;
+        for (;;) {
+            // Snapshot wake_ BEFORE reading tail_ to prevent lost wakeups.
+            const uint32_t    w = wake_.load(std::memory_order_relaxed);
+            const std::size_t h = head_.load(std::memory_order_relaxed);
+                  std::size_t t = tail_.load(std::memory_order_acquire);
+
+            // If empty, spin before sleeping: avoids the futex when the next item
+            // arrives within the spin window (~4 µs at default spin_count=200 on x86).
+            if (h == t) {
+                if (!accepting_.load(std::memory_order_acquire))
+                    throw ChannelClosedError{};
+
+                for (std::size_t s = 0; s < spin_count_; ++s) {
+                    spin_hint();
+                    t = tail_.load(std::memory_order_acquire);
+                    if (t != h) break;
+                    if (!accepting_.load(std::memory_order_relaxed))
+                        throw ChannelClosedError{};
+                }
+
+                if (h == t) {
+                    // Still empty after spin — sleep until push() or disable() fires.
+                    // Re-check tail after loading w to guard against a lost wakeup.
+                    if (tail_.load(std::memory_order_acquire) != h) continue;
+                    wake_.wait(w, std::memory_order_relaxed);
+                    continue;
+                }
+            }
+
+            // Item available (found immediately or during spin).
+            if (!accepting_.load(std::memory_order_acquire))
+                throw ChannelClosedError{};
+            T value = extract(std::move(buf_[h & ring_mask_]));
+            head_.store(h + 1, std::memory_order_release);
+            stats_.record_pop();
+            return value;
+        }
     }
 
-    // Non-blocking pop with timeout. For watchdog/display use only — not used in run_loop.
+    // Non-blocking pop with timeout. For watchdog/display use only.
     bool try_pop(T& out, std::chrono::milliseconds timeout) {
-        std::unique_lock lock(mutex_);
-        if (!cv_.wait_for(lock, timeout, [this] {
-                return !queue_.empty() || !accepting_.load(std::memory_order_relaxed);
-            }))
-            return false;
-        if (queue_.empty())
-            return false;
-        out = extract(std::move(queue_.front()));
-        queue_.pop();
+        const auto deadline = std::chrono::steady_clock::now() + timeout;
+        for (;;) {
+            if (try_pop_now(out)) return true;
+            if (!accepting_.load(std::memory_order_relaxed)) return false;
+            if (std::chrono::steady_clock::now() >= deadline) return false;
+            std::this_thread::sleep_for(std::chrono::microseconds(50));
+        }
+    }
+
+    // Immediate non-blocking pop. Returns false if the ring is empty.
+    bool try_pop_now(T& out) {
+        const std::size_t h = head_.load(std::memory_order_relaxed);
+        if (h == tail_.load(std::memory_order_acquire)) return false;
+        out = extract(std::move(buf_[h & ring_mask_]));
+        head_.store(h + 1, std::memory_order_release);
         stats_.record_pop();
         return true;
     }
@@ -114,38 +204,44 @@ public:
         accepting_.store(true, std::memory_order_relaxed);
     }
 
-    // Disable the channel: drop all queued items, unblock any waiting pop().
-    // Called by consumer node on stop(). Producer push() will silently drop after this.
+    // Disable the channel: stop accepting new pushes, unblock any waiting pop().
+    // Items already in the ring are abandoned and freed when the Channel is destroyed.
     void disable() {
-        accepting_.store(false, std::memory_order_relaxed);
-        {
-            std::lock_guard lock(mutex_);
-            while (!queue_.empty()) queue_.pop();
-        }
-        cv_.notify_all();
+        accepting_.store(false, std::memory_order_release);
+        wake_.fetch_add(1, std::memory_order_release);
+        wake_.notify_all();
     }
 
-    std::size_t size() const {
-        std::lock_guard lock(mutex_);
-        return queue_.size();
+    // Register a callback fired when the queue transitions empty→non-empty.
+    void set_push_callback(std::function<void()> cb) {
+        push_callback_ = std::move(cb);
     }
 
+    // Size derived lazily from ring indices — no separate counter on the hot path.
+    std::size_t size() const {
+        return tail_.load(std::memory_order_relaxed)
+             - head_.load(std::memory_order_relaxed);
+    }
+    std::size_t approx_size() const { return size(); }
+
     std::size_t capacity()     const { return capacity_; }
     bool        is_accepting() const { return accepting_.load(std::memory_order_relaxed); }
     const ChannelStats& stats() const { return stats_; }
 
     ChannelSnapshot snapshot(const std::string& name) const {
-        std::lock_guard lock(mutex_);
+        const std::size_t t = tail_.load(std::memory_order_relaxed);
+        const std::size_t h = head_.load(std::memory_order_relaxed);
         return {
             name,
             capacity_,
-            queue_.size(),
+            t - h,
             stats_.peak_fill.load(std::memory_order_relaxed),
             stats_.pushes.load(std::memory_order_relaxed),
+            stats_.bytes_pushed.load(std::memory_order_relaxed),
             stats_.drops.load(std::memory_order_relaxed),
             stats_.overflows.load(std::memory_order_relaxed),
             stats_.pops.load(std::memory_order_relaxed),
-            sizeof(T),   // payload bytes — sizeof(T) regardless of storage policy
+            sizeof(T),
         };
     }
 
@@ -164,12 +260,19 @@ private:
             return *s;
     }
 
-    const std::size_t            capacity_;
-    std::queue<storage_type>     queue_;
-    std::atomic<bool>            accepting_{true};
-    mutable std::mutex           mutex_;
-    std::condition_variable      cv_;
-    ChannelStats                 stats_;
+    const std::size_t               capacity_;
+    const std::size_t               spin_count_;
+    std::size_t                     ring_mask_;
+    std::unique_ptr<storage_type[]> buf_;
+    std::function<void()>           push_callback_;
+    ChannelStats                    stats_;
+
+    // Separate cache lines: head_ is written only by the consumer;
+    // tail_ and wake_ are written only by the producer.
+    alignas(64) std::atomic<std::size_t> head_{0};
+    alignas(64) std::atomic<std::size_t> tail_{0};
+                std::atomic<uint32_t>    wake_{0};
+                std::atomic<bool>        accepting_{true};
 };
 
 // ── Channel probe — type-erased snapshot accessor ─────────────────────────────

include/kpn/debug_hub.hpp

@@ -0,0 +1,456 @@
+#pragma once
+// Only active when KPN_WEB_DEBUG is defined.
+
+#ifdef KPN_WEB_DEBUG
+#include "diagnostics.hpp"
+#include "web_debug.hpp"
+
+#include <functional>
+#include <memory>
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
+
+namespace kpn {
+
+// ── Hub HTML ──────────────────────────────────────────────────────────────────
+// Multi-tab UI: one tab per registered network (force-directed graph) +
+// an "All Networks" tab showing shared resource cards and a cross-network
+// node table.
+
+static const char* HUB_HTML = R"html(<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<title>KPN++ Debug Hub</title>
+<style>
+  *{box-sizing:border-box}
+  body{margin:0;background:#1a1a2e;color:#eee;font-family:monospace}
+  #hdr{display:flex;align-items:center;padding:0 16px;background:#16213e;
+       border-bottom:1px solid #0f3460;height:44px;gap:8px;overflow-x:auto}
+  #hdr h1{margin:0;font-size:16px;color:#e94560;white-space:nowrap;margin-right:12px}
+  #tab-bar{display:flex;gap:2px;flex:1}
+  .tab{padding:4px 14px;border:none;background:#0f3460;color:#aaa;
+       cursor:pointer;font-family:monospace;font-size:11px;border-radius:2px;white-space:nowrap}
+  .tab.active{background:#e94560;color:#fff}
+  .tab:hover:not(.active){background:#1e3a6e;color:#eee}
+  #status{font-size:10px;color:#555;white-space:nowrap}
+
+  .panel{display:none}
+  .panel.active{display:block}
+
+  /* ── All Networks tab ─────────────────────────────────────── */
+  #panel-all{height:calc(100vh - 44px);overflow-y:auto;padding:16px;
+             display:none;gap:16px;grid-template-columns:300px 1fr}
+  #panel-all.active{display:grid;align-content:start}
+  #panel-all h2{font-size:11px;color:#e94560;margin:0 0 8px;
+                letter-spacing:1px;text-transform:uppercase}
+
+  #res-col{grid-column:1}
+  .res-card{background:#16213e;border:1px solid #0f3460;border-radius:4px;
+            padding:10px 12px;margin-bottom:8px}
+  .res-head{display:flex;justify-content:space-between;align-items:center}
+  .res-name{font-size:12px}
+  .badge{font-size:9px;padding:2px 6px;border-radius:2px}
+  .held{background:#e94560}.free{background:#4CAF50;color:#111}
+  .res-meta{font-size:9px;color:#666;margin-top:5px;display:flex;gap:12px;flex-wrap:wrap}
+  .bar-wrap{height:3px;background:#0f3460;border-radius:2px;margin-top:7px}
+  .bar{height:3px;border-radius:2px;transition:width 0.4s}
+
+  #nodes-col{grid-column:2;overflow-y:auto;max-height:calc(100vh - 76px)}
+  table{width:100%;border-collapse:collapse;font-size:10px}
+  th{padding:4px 8px;color:#555;border-bottom:1px solid #0f3460;text-align:left;
+     position:sticky;top:0;background:#1a1a2e;z-index:1}
+  td{padding:2px 8px;border-bottom:1px solid #16213e}
+  tr:hover td{background:#16213e}
+  .ntag{font-size:9px;background:#0f3460;padding:1px 4px;border-radius:2px;color:#4CAF50}
+
+  /* ── Per-network graph panels ─────────────────────────────── */
+  .graph-panel{width:100vw;height:calc(100vh - 44px)}
+  svg.net{width:100%;height:100%}
+  .node circle{stroke:#fff;stroke-width:1.5px}
+  .node text{font-size:11px;fill:#eee;pointer-events:none;text-anchor:middle}
+  .node .st{font-size:9px;fill:#aaa}
+  .link{fill:none;stroke-width:2px}
+  .lbl{font-size:9px;fill:#ccc}
+  #tip{position:absolute;background:#0f3460;border:1px solid #e94560;border-radius:4px;
+       padding:8px 12px;font-size:11px;pointer-events:none;display:none;
+       white-space:pre;line-height:1.6}
+</style>
+</head>
+<body>
+<div id="hdr">
+  <h1>KPN++ Debug Hub</h1>
+  <div id="tab-bar"></div>
+  <span id="status">connecting…</span>
+</div>
+<div id="panels">
+  <div id="panel-all" class="panel"></div>
+</div>
+<div id="tip"></div>
+<script src="https://d3js.org/d3.v7.min.js"></script>
+<script>
+const R = 28;
+const tip = d3.select('#tip');
+
+const nc = ema  => ema>100?'#e94560':ema>50?'#e07040':ema>10?'#f0c040':'#4CAF50';
+const ec = pct  => pct>=80?'#e94560':pct>=50?'#f0c040':'#4CAF50';
+const ea = pct  => pct>=80?'url(#a2)':pct>=50?'url(#a1)':'url(#a0)';
+
+// ── Tab management ────────────────────────────────────────────────────────────
+let activeTab = null;
+
+function ensureTab(id, label) {
+  if (document.getElementById('tab-' + id)) return;
+  const b = document.createElement('button');
+  b.className = 'tab'; b.id = 'tab-' + id; b.textContent = label;
+  b.onclick = () => showTab(id);
+  document.getElementById('tab-bar').appendChild(b);
+}
+
+function showTab(id) {
+  activeTab = id;
+  document.querySelectorAll('.tab').forEach(b =>
+    b.classList.toggle('active', b.id === 'tab-' + id));
+  document.querySelectorAll('.panel').forEach(p =>
+    p.classList.toggle('active', p.id === 'panel-' + id));
+}
+
+// ── All Networks tab ──────────────────────────────────────────────────────────
+function renderAll(data) {
+  const panel = document.getElementById('panel-all');
+
+  // Resources column
+  let rhtml = '<div id="res-col"><h2>Shared Resources</h2>';
+  if (!data.resources || !data.resources.length)
+    rhtml += '<div style="color:#444;font-size:11px">None registered</div>';
+  for (const r of (data.resources || [])) {
+    const wpct = Math.min(100, r.avg_wait_ms).toFixed(1);
+    const bc   = r.current_waiters > 0 ? '#e94560' : '#4CAF50';
+    rhtml += `<div class="res-card">
+      <div class="res-head">
+        <span class="res-name">${r.name}</span>
+        <span class="badge ${r.held ? 'held' : 'free'}">${r.held ? 'HELD' : 'free'}</span>
+      </div>
+      <div class="res-meta">
+        <span>avg wait ${r.avg_wait_ms.toFixed(1)} ms</span>
+        <span>waiters ${r.current_waiters} / peak ${r.peak_waiters}</span>
+        <span>${r.acquisitions} acq</span>
+      </div>
+      <div class="bar-wrap">
+        <div class="bar" style="width:${wpct}%;background:${bc}"></div>
+      </div>
+    </div>`;
+  }
+  rhtml += '</div>';
+
+  // Nodes column — all networks in one table
+  let rows = '';
+  for (const net of data.networks) {
+    for (const n of net.nodes) {
+      rows += `<tr>
+        <td><span class="ntag">${net.name}</span></td>
+        <td>${n.id}</td>
+        <td>${n.fps.toFixed(1)}</td>
+        <td>${n.ema_exec_ms.toFixed(2)}</td>
+        <td>${n.max_exec_ms.toFixed(2)}</td>
+        <td>${n.blocked_ms.toFixed(2)}</td>
+        <td>${n.cpu_util_pct.toFixed(1)}</td>
+      </tr>`;
+    }
+  }
+  const thtml = `<div id="nodes-col"><h2>All Nodes</h2>
+    <table>
+      <tr><th>Network</th><th>Node</th><th>fps</th>
+          <th>exec ema (ms)</th><th>exec max (ms)</th>
+          <th>blocked (ms)</th><th>cpu %</th></tr>
+      ${rows}
+    </table></div>`;
+
+  panel.innerHTML = rhtml + thtml;
+}
+
+// ── Per-network graph ─────────────────────────────────────────────────────────
+const nets = {};
+
+function initNet(netData) {
+  const name = netData.name;
+
+  const div = document.createElement('div');
+  div.id = 'panel-' + name;
+  div.className = 'panel graph-panel';
+  document.getElementById('panels').appendChild(div);
+
+  const svg = d3.select(div).append('svg').attr('class', 'net');
+  const W = () => div.clientWidth  || window.innerWidth;
+  const H = () => div.clientHeight || (window.innerHeight - 44);
+
+  const defs = svg.append('defs');
+  ['#4CAF50','#f0c040','#e94560'].forEach((col, i) =>
+    defs.append('marker').attr('id','a'+i)
+      .attr('viewBox','0 -5 10 10').attr('refX',10).attr('refY',0)
+      .attr('markerWidth',6).attr('markerHeight',6).attr('orient','auto')
+      .append('path').attr('d','M0,-5L10,0L0,5').attr('fill',col));
+
+  const g = svg.append('g');
+  svg.call(d3.zoom().on('zoom', e => g.attr('transform', e.transform)));
+
+  const nodes = netData.nodes.map(n => ({...n, x: W()/2, y: H()/2}));
+  const byId  = Object.fromEntries(nodes.map(n => [n.id, n]));
+  const links = netData.edges
+    .map(e => ({...e, source: byId[e.source], target: byId[e.target]}))
+    .filter(e => e.source && e.target);
+
+  const sim = d3.forceSimulation(nodes)
+    .force('link',    d3.forceLink(links).distance(150).strength(0.5))
+    .force('charge',  d3.forceManyBody().strength(-350))
+    .force('center',  d3.forceCenter(W()/2, H()/2))
+    .force('collide', d3.forceCollide(R + 18))
+    .on('tick', tick);
+
+  const lsel = g.append('g').selectAll('line').data(links).join('line')
+    .attr('class','link')
+    .attr('stroke', d => ec(d.fill_pct))
+    .attr('marker-end', d => ea(d.fill_pct));
+
+  const llbl = g.append('g').selectAll('text').data(links).join('text')
+    .attr('class','lbl').text(d => `${d.fill_pct.toFixed(0)}%`);
+
+  const ng = g.append('g').selectAll('g').data(nodes).join('g').attr('class','node')
+    .call(d3.drag()
+      .on('start',(e,d)=>{ if(!e.active) sim.alphaTarget(0.3).restart(); d.fx=d.x; d.fy=d.y; })
+      .on('drag', (e,d)=>{ d.fx=e.x; d.fy=e.y; })
+      .on('end',  (e,d)=>{ if(!e.active) sim.alphaTarget(0); d.fx=null; d.fy=null; }));
+
+  ng.append('circle').attr('r', R).attr('fill', d => nc(d.ema_exec_ms));
+  ng.append('text').attr('dy', 4).text(d => d.id);
+  ng.append('text').attr('class','st').attr('dy', 18)
+    .text(d => `${d.ema_exec_ms.toFixed(1)}ms ${d.fps.toFixed(1)}fps`);
+
+  ng.on('mousemove', (e,d) =>
+      tip.style('display','block')
+         .style('left',(e.pageX+12)+'px').style('top',(e.pageY+12)+'px')
+         .text(`${d.id}\nframes: ${d.frames}  fps: ${d.fps.toFixed(2)}\n` +
+               `exec ema: ${d.ema_exec_ms.toFixed(2)}ms  max: ${d.max_exec_ms.toFixed(2)}ms\n` +
+               `blocked: ${d.blocked_ms.toFixed(2)}ms  cpu: ${d.cpu_util_pct.toFixed(1)}%`))
+    .on('mouseleave', () => tip.style('display','none'));
+
+  g.selectAll('.link')
+    .on('mousemove', (e,d) =>
+      tip.style('display','block')
+         .style('left',(e.pageX+12)+'px').style('top',(e.pageY+12)+'px')
+         .text(`${d.name}\nfill: ${d.fill_pct.toFixed(1)}%  peak: ${d.peak_pct.toFixed(1)}%\n` +
+               `cap: ${d.capacity}  pushes: ${d.pushes}  drops: ${d.drops}\n` +
+               `bandwidth: ${(d.bw_mbs||0).toFixed(2)} MB/s`))
+    .on('mouseleave', () => tip.style('display','none'));
+
+  function tick() {
+    const w = W(), h = H();
+    nodes.forEach(d => {
+      d.x = Math.max(R, Math.min(w - R, d.x));
+      d.y = Math.max(R, Math.min(h - R, d.y));
+    });
+    lsel
+      .attr('x1', d => d.source.x).attr('y1', d => d.source.y)
+      .attr('x2', d => { const dx=d.target.x-d.source.x, dy=d.target.y-d.source.y,
+                               dist=Math.sqrt(dx*dx+dy*dy)||1;
+                         return d.target.x-(dx/dist)*(R+8); })
+      .attr('y2', d => { const dx=d.target.x-d.source.x, dy=d.target.y-d.source.y,
+                               dist=Math.sqrt(dx*dx+dy*dy)||1;
+                         return d.target.y-(dy/dist)*(R+8); });
+    llbl.attr('x', d => (d.source.x+d.target.x)/2)
+        .attr('y', d => (d.source.y+d.target.y)/2 - 6);
+    ng.attr('transform', d => `translate(${d.x},${d.y})`);
+  }
+
+  nets[name] = {nodes, links, sim, ng, lsel, llbl};
+}
+
+function updateNet(netData) {
+  const st = nets[netData.name];
+  if (!st) return;
+  const byId = Object.fromEntries(netData.nodes.map(n => [n.id, n]));
+  st.nodes.forEach(n => {
+    const f = byId[n.id];
+    if (f) Object.assign(n, {frames:f.frames, ema_exec_ms:f.ema_exec_ms,
+      max_exec_ms:f.max_exec_ms, blocked_ms:f.blocked_ms, fps:f.fps,
+      total_cpu_ms:f.total_cpu_ms, cpu_util_pct:f.cpu_util_pct});
+  });
+  netData.edges.forEach((e,i) => {
+    if (st.links[i]) Object.assign(st.links[i], {fill_pct:e.fill_pct,
+      peak_pct:e.peak_pct, pushes:e.pushes, drops:e.drops,
+      overflows:e.overflows, current:e.current, bw_mbs:e.bw_mbs});
+  });
+  st.ng.select('circle').attr('fill', d => nc(d.ema_exec_ms));
+  st.ng.select('.st').text(d => `${d.ema_exec_ms.toFixed(1)}ms ${d.fps.toFixed(1)}fps`);
+  st.lsel.attr('stroke', d => ec(d.fill_pct)).attr('marker-end', d => ea(d.fill_pct));
+  st.llbl.text(d => `${d.fill_pct.toFixed(0)}%`);
+}
+
+// ── Poll loop ─────────────────────────────────────────────────────────────────
+let init = false;
+
+async function poll() {
+  try {
+    const r = await fetch('/api/snapshot');
+    if (!r.ok) throw new Error(r.status);
+    const data = await r.json();
+
+    if (!init) {
+      ensureTab('all', 'All Networks');
+      data.networks.forEach(net => ensureTab(net.name, net.name));
+      showTab('all');
+      data.networks.forEach(initNet);
+      init = true;
+    }
+
+    renderAll(data);
+    data.networks.forEach(updateNet);
+
+    document.getElementById('status').textContent =
+      `${new Date().toLocaleTimeString()} · ${data.networks.length} nets · ${(data.resources||[]).length} resources`;
+  } catch(e) {
+    document.getElementById('status').textContent = 'error: ' + e;
+  }
+}
+
+poll();
+setInterval(poll, 500);
+window.addEventListener('resize', () =>
+  Object.values(nets).forEach(st =>
+    st.sim.force('center', d3.forceCenter(
+      (document.getElementById('panel-' + Object.keys(nets).find(k => nets[k] === st))?.clientWidth  || window.innerWidth)  / 2,
+      (document.getElementById('panel-' + Object.keys(nets).find(k => nets[k] === st))?.clientHeight || window.innerHeight - 44) / 2
+    )).alpha(0.1).restart()));
+</script>
+</body>
+</html>
+)html";
+
+// ── DebugHub ──────────────────────────────────────────────────────────────────
+
+class DebugHub {
+public:
+    explicit DebugHub(uint16_t port = 9090) : port_(port) {}
+
+    ~DebugHub() { stop(); }
+    DebugHub(const DebugHub&)            = delete;
+    DebugHub& operator=(const DebugHub&) = delete;
+
+    // Register a network. Disables that network's own web server so the hub
+    // becomes the single debug endpoint. Call before network.start().
+    template<typename Net>
+    void register_network(const std::string& name, Net& net) {
+        net.disable_web_server();
+        networks_.push_back({name, [&net, name] {
+            auto s = net.network_snapshot();
+            s.name = name;
+            return s;
+        }});
+    }
+
+    // Register a shared resource — appears in the "All Networks" resource panel.
+    void register_resource(const std::string& name, IResourceProbe* probe) {
+        resources_.emplace_back(name, probe);
+    }
+
+    void start() {
+        server_ = std::make_unique<web_debug::WebDebugServer>(
+            port_,
+            [this] { return build_json(); },
+            HUB_HTML);
+        server_->start();
+        std::cerr << "[kpn] hub debug UI: http://localhost:" << port_ << "\n";
+    }
+
+    void stop() { if (server_) server_->stop(); }
+
+private:
+    // Serialise nodes array for one network snapshot
+    static void write_nodes(std::ostream& o, const NetworkSnapshot& s) {
+        o << "[";
+        for (std::size_t i = 0; i < s.nodes.size(); ++i) {
+            const auto& n = s.nodes[i];
+            if (i) o << ',';
+            o << "{\"id\":\""          << web_debug::escape_json(n.name) << "\""
+              << ",\"frames\":"        << n.frames_processed
+              << ",\"ema_exec_ms\":"   << n.ema_exec_ms
+              << ",\"max_exec_ms\":"   << n.max_exec_ms
+              << ",\"blocked_ms\":"    << n.total_blocked_ms
+              << ",\"fps\":"           << n.throughput_fps
+              << ",\"total_cpu_ms\":"  << n.total_cpu_ms
+              << ",\"cpu_util_pct\":"  << n.cpu_util_pct
+              << "}";
+        }
+        o << "]";
+    }
+
+    // Serialise edges array for one network snapshot
+    static void write_edges(std::ostream& o, const NetworkSnapshot& s) {
+        o << "[";
+        for (std::size_t i = 0; i < s.channels.size(); ++i) {
+            const auto& c = s.channels[i];
+            if (i) o << ',';
+            auto [src, dst] = web_debug::parse_edge_name(c.name);
+            o << "{\"name\":\""    << web_debug::escape_json(c.name) << "\""
+              << ",\"source\":\""  << web_debug::escape_json(src)    << "\""
+              << ",\"target\":\""  << web_debug::escape_json(dst)    << "\""
+              << ",\"capacity\":"  << c.capacity
+              << ",\"current\":"   << c.current_fill
+              << ",\"fill_pct\":"  << c.fill_pct()
+              << ",\"peak_pct\":"  << c.peak_pct()
+              << ",\"pushes\":"    << c.pushes
+              << ",\"drops\":"     << c.drops
+              << ",\"overflows\":" << c.overflows
+              << ",\"item_bytes\":" << c.item_bytes
+              << ",\"bw_mbs\":"    << c.bandwidth_mbs(s.elapsed_s)
+              << "}";
+        }
+        o << "]";
+    }
+
+    std::string build_json() const {
+        std::ostringstream o;
+        o << std::fixed;
+        o.precision(2);
+
+        o << "{\"networks\":[";
+        for (std::size_t i = 0; i < networks_.size(); ++i) {
+            if (i) o << ',';
+            auto s = networks_[i].fn();
+            o << "{\"name\":\"" << web_debug::escape_json(networks_[i].name) << "\""
+              << ",\"nodes\":";  write_nodes(o, s);
+            o << ",\"edges\":";  write_edges(o, s);
+            o << "}";
+        }
+
+        o << "],\"resources\":[";
+        for (std::size_t i = 0; i < resources_.size(); ++i) {
+            if (i) o << ',';
+            const auto r = resources_[i].second->snapshot(resources_[i].first);
+            o << "{\"name\":\""          << web_debug::escape_json(r.name) << "\""
+              << ",\"acquisitions\":"    << r.acquisitions
+              << ",\"avg_wait_ms\":"     << r.avg_wait_ms
+              << ",\"peak_waiters\":"    << r.peak_waiters
+              << ",\"current_waiters\":" << r.current_waiters
+              << ",\"held\":"            << (r.held ? "true" : "false")
+              << "}";
+        }
+        o << "]}";
+        return o.str();
+    }
+
+    struct Entry {
+        std::string                      name;
+        std::function<NetworkSnapshot()> fn;
+    };
+
+    uint16_t                                         port_;
+    std::vector<Entry>                               networks_;
+    std::vector<std::pair<std::string,IResourceProbe*>> resources_;
+    std::unique_ptr<web_debug::WebDebugServer>       server_;
+};
+
+} // namespace kpn
+#endif // KPN_WEB_DEBUG

include/kpn/diagnostics.hpp

@@ -15,6 +15,7 @@ using duration_t = std::chrono::duration<double, std::milli>; // milliseconds
 
 struct ChannelStats {
     std::atomic<uint64_t> pushes{0};
+    std::atomic<uint64_t> bytes_pushed{0};
     std::atomic<uint64_t> drops{0};
     std::atomic<uint64_t> overflows{0};
     std::atomic<uint64_t> pops{0};
@@ -24,8 +25,9 @@ struct ChannelStats {
     ChannelStats(const ChannelStats&) = delete;
     ChannelStats& operator=(const ChannelStats&) = delete;
 
-    void record_push(std::size_t current_fill) {
+    void record_push(std::size_t current_fill, std::size_t data_bytes) {
         pushes.fetch_add(1, std::memory_order_relaxed);
+        bytes_pushed.fetch_add(data_bytes, std::memory_order_relaxed);
         std::size_t prev = peak_fill.load(std::memory_order_relaxed);
         while (current_fill > prev &&
                !peak_fill.compare_exchange_weak(prev, current_fill,
@@ -54,6 +56,12 @@ struct NodeStats {
     // blocked on mutexes/channels. Sampled once per frame.
     std::atomic<int64_t> total_cpu_us{0};  // cumulative CPU µs consumed
 
+    // Pool scheduling stats — only meaningful for PoolNode / InterruptNode.
+    // exec_start_us: wall-clock µs when fire_once began; 0 when idle.
+    // Used by the watchdog to detect hung nodes (elapsed > max_exec_time).
+    std::atomic<int64_t> queue_wait_us{0};  // cumulative µs spent in pool queue
+    std::atomic<int64_t> exec_start_us{0};  // non-zero while fire_once is running
+
     NodeStats() = default;
     NodeStats(const NodeStats&) = delete;
     NodeStats& operator=(const NodeStats&) = delete;
@@ -71,6 +79,11 @@ struct NodeStats {
              + static_cast<int64_t>(ts.tv_nsec) / 1'000;
     }
 
+    void record_queue_wait(duration_t wait) {
+        int64_t us = static_cast<int64_t>(wait.count() * 1000.0);
+        if (us > 0) queue_wait_us.fetch_add(us, std::memory_order_relaxed);
+    }
+
     void record_exec(duration_t exec_time, duration_t blocked_time,
                      const struct timespec& cpu_before, const struct timespec& cpu_after) {
         frames_processed.fetch_add(1, std::memory_order_relaxed);
@@ -105,10 +118,11 @@ struct ChannelSnapshot {
     std::size_t current_fill;
     std::size_t peak_fill;
     uint64_t    pushes;
+    uint64_t    bytes_pushed;  // actual bytes accumulated via channel_data_size<T>
     uint64_t    drops;
     uint64_t    overflows;
     uint64_t    pops;
-    std::size_t item_bytes;   // sizeof(T) for the stored type — set by Channel<T>
+    std::size_t item_bytes;   // sizeof(T) — nominal struct size, not necessarily data size
 
     double fill_pct() const {
         return capacity ? 100.0 * current_fill / capacity : 0.0;
@@ -116,10 +130,10 @@ struct ChannelSnapshot {
     double peak_pct() const {
         return capacity ? 100.0 * peak_fill / capacity : 0.0;
     }
-    // Bandwidth in MB/s: bytes transferred / elapsed seconds
+    // Bandwidth in MB/s: actual bytes transferred / elapsed seconds
     double bandwidth_mbs(double elapsed_s) const {
-        if (elapsed_s <= 0.0 || item_bytes == 0) return 0.0;
-        return static_cast<double>(pushes * item_bytes) / elapsed_s / 1e6;
+        if (elapsed_s <= 0.0) return 0.0;
+        return static_cast<double>(bytes_pushed) / elapsed_s / 1e6;
     }
 };
 
@@ -128,10 +142,52 @@ struct NodeSnapshot {
     uint64_t    frames_processed;
     double      ema_exec_ms;
     double      max_exec_ms;
-    double      total_blocked_ms;
+    double      total_blocked_ms;   // ThreadPerNode: time blocked in channel pop
     double      throughput_fps;
-    double      total_cpu_ms;     // cumulative CPU time consumed by this node's thread
-    double      cpu_util_pct;     // exec_ms / (exec_ms + blocked_ms) * 100
+    double      total_cpu_ms;       // cumulative CPU time consumed by this node's thread
+    double      cpu_util_pct;       // exec_ms / (exec_ms + blocked_ms) * 100
+    double      queue_wait_ms{0};   // PoolNode: cumulative time spent in pool queue
+};
+
+// ── Pool statistics + snapshot ────────────────────────────────────────────────
+
+struct PoolSnapshot {
+    std::string  name;
+    std::size_t  thread_count;
+    std::size_t  queue_depth;     // tasks waiting in the priority queue
+    std::size_t  active_count;    // tasks currently executing
+    uint64_t     tasks_submitted;
+    uint64_t     tasks_completed;
+};
+
+struct IPoolProbe {
+    virtual ~IPoolProbe() = default;
+    virtual PoolSnapshot snapshot(const std::string& name) const = 0;
+};
+
+// ── Cross-network snapshot (used by DebugHub) ─────────────────────────────────
+
+struct NetworkSnapshot {
+    std::string                  name;
+    std::vector<NodeSnapshot>    nodes;
+    std::vector<ChannelSnapshot> channels;
+    double                       elapsed_s;
+};
+
+// ── Resource statistics + snapshot ───────────────────────────────────────────
+
+struct ResourceSnapshot {
+    std::string name;
+    uint64_t    acquisitions;
+    double      avg_wait_ms;
+    uint64_t    peak_waiters;
+    uint64_t    current_waiters;
+    bool        held;
+};
+
+struct IResourceProbe {
+    virtual ~IResourceProbe() = default;
+    virtual ResourceSnapshot snapshot(const std::string& name) const = 0;
 };
 
 } // namespace kpn

include/kpn/fanout.hpp

@@ -1,7 +1,7 @@
 #pragma once
 #include "channel.hpp"
 #include "diagnostics.hpp"
-#include "node.hpp"
+#include "inode.hpp"
 #include "port.hpp"
 #include "traits.hpp"
 
@@ -15,24 +15,6 @@
 
 namespace kpn {
 
-namespace detail {
-
-// Produces std::tuple<T, T, ..., T> with N repetitions — used so that
-// Network::connect can do its normal return_tuple type-check against FanoutNode.
-template<typename T, std::size_t N, typename Seq = std::make_index_sequence<N>>
-struct repeat_tuple;
-
-template<typename T, std::size_t N, std::size_t... Is>
-struct repeat_tuple<T, N, std::index_sequence<Is...>> {
-    template<std::size_t> using always_T = T;
-    using type = std::tuple<always_T<Is>...>;
-};
-
-template<typename T, std::size_t N>
-using repeat_tuple_t = typename repeat_tuple<T, N>::type;
-
-} // namespace detail
-
 // ── FanoutNode ────────────────────────────────────────────────────────────────
 //
 // Reads one item from its single input channel and pushes a copy to each of
@@ -48,7 +30,7 @@ template<typename T, std::size_t N, std::size_t Id = 0>
 class FanoutNode : public INode {
 public:
     using args_tuple   = std::tuple<T>;
-    using return_tuple = detail::repeat_tuple_t<T, N>;
+    using return_tuple = repeat_tuple_t<T, N>;
     using return_raw   = return_tuple;
 
     static constexpr std::size_t input_count    = 1;

include/kpn/inode.hpp

@@ -0,0 +1,45 @@
+#pragma once
+#include "diagnostics.hpp"
+#include <chrono>
+#include <functional>
+#include <string>
+#include <string_view>
+
+namespace kpn {
+
+// Called when a node's function throws. Return true to skip the failed
+// invocation and keep running, false to stop the node.
+using NodeErrorHandler = std::function<bool(std::string_view node_name, std::exception_ptr)>;
+
+// Lightweight timestamp-only callback fired on per-node events.
+// The node name is known at registration time so it is not included here.
+using NodeEventCallback = std::function<void(std::chrono::steady_clock::time_point)>;
+
+// Event types reported to the network-level aggregate callback.
+enum class NodeEvent { Overflow, Closed };
+
+// ── INode — type-erased interface for Network / watchdog ─────────────────────
+
+struct INode {
+    virtual ~INode()       = default;
+    virtual void start()   = 0;
+    virtual void stop()    = 0;
+    virtual bool running() const = 0;
+    virtual const NodeStats& stats() const = 0;
+    virtual NodeSnapshot node_snapshot(const std::string& name, double elapsed_s) const = 0;
+    virtual void set_name(std::string name) = 0;
+
+    // Network-injected callbacks (slot 1 of each node's callback array).
+    // Default no-ops; overridden by PoolNode, PoolObjectNode, InterruptNode.
+    virtual void set_network_overflow_callback(NodeEventCallback) {}
+    virtual void set_network_closed_callback(NodeEventCallback)   {}
+
+    // halt(): alias for stop() — immediate, discards in-flight work.
+    virtual void halt() { stop(); }
+
+    // shutdown(): graceful drain before stopping. Base implementation falls
+    // back to stop(). Network and StaticNetwork override with topo-ordered drain.
+    virtual void shutdown() { stop(); }
+};
+
+} // namespace kpn

include/kpn/interrupt_node.hpp

@@ -0,0 +1,279 @@
+#pragma once
+#include "channel.hpp"
+#include "diagnostics.hpp"
+#include "fixed_string.hpp"
+#include "inode.hpp"
+#include "port.hpp"
+#include "scheduler.hpp"
+#include "traits.hpp"
+
+#include <array>
+#include <atomic>
+#include <chrono>
+#include <cstddef>
+#include <functional>
+#include <iostream>
+#include <memory>
+#include <string>
+#include <tuple>
+#include <type_traits>
+
+namespace kpn {
+
+// ── InterruptNode ─────────────────────────────────────────────────────────────
+//
+// A source node (zero inputs) driven by an external event — camera frame ready,
+// timer tick, socket data, etc. — rather than self-submission.
+//
+// Usage:
+//   auto node = make_interrupt_node<produce_frame>(pool, out<"frame">{});
+//   camera_sdk.on_frame_ready(node.get_trigger());  // register with external source
+//   network.add("camera", node).connect(...).build().start();
+//
+// The trigger callable is safe to call from any thread, including signal handlers,
+// provided the underlying scheduler's submit() is signal-safe. After fire_once()
+// completes the node is idle until the next trigger fires — it does NOT busy-loop.
+
+template<auto Func,
+         typename OutputTag = out<>,
+         fixed_string Label    = "",
+         std::size_t UniqueTag = 0>
+class InterruptNode;
+
+template<auto Func, fixed_string... OutNames, fixed_string Label, std::size_t UniqueTag>
+class InterruptNode<Func, out<OutNames...>, Label, UniqueTag> : public INode {
+public:
+    using F            = decltype(Func);
+    using return_raw   = return_t<F>;
+    using return_tuple = normalised_return_t<return_raw>;
+
+    static_assert(arity_v<F> == 0,
+        "InterruptNode function must take no arguments (it has no input channels)");
+
+    static constexpr std::string_view label()     { return Label.view(); }
+    static constexpr std::size_t      unique_tag  = UniqueTag;
+    static constexpr std::size_t      input_count  = 0;
+    static constexpr std::size_t      output_count = std::tuple_size_v<return_tuple>;
+
+    static_assert(
+        sizeof...(OutNames) == 0 || sizeof...(OutNames) == output_count,
+        "make_interrupt_node: number of output names must match return tuple size, or provide none"
+    );
+
+    explicit InterruptNode(std::shared_ptr<IScheduler> sched, std::size_t fifo_capacity = 5)
+        : scheduler_(std::move(sched)), fifo_capacity_(fifo_capacity)
+    {}
+
+    ~InterruptNode() override { stop(); }
+
+    // ── INode ─────────────────────────────────────────────────────────────────
+
+    void start() override {
+        stop_flag_.store(false, std::memory_order_relaxed);
+        pending_.store(0, std::memory_order_relaxed);
+        // Does NOT self-submit — waits for first external trigger.
+    }
+
+    void stop() override {
+        stop_flag_.store(true, std::memory_order_seq_cst);
+        // In-flight fire_once() observes stop_flag_ on its next check.
+    }
+
+    bool running() const override { return !stop_flag_.load(std::memory_order_relaxed); }
+    void set_name(std::string name) override { name_ = std::move(name); }
+    void set_error_handler(NodeErrorHandler h) { error_handler_ = std::move(h); }
+    void set_max_exec_time(std::chrono::milliseconds t) { max_exec_time_ = t; }
+
+    void set_overflow_callback(NodeEventCallback cb) { event_callbacks_[0] = std::move(cb); }
+    void set_network_overflow_callback(NodeEventCallback cb) override { event_callbacks_[1] = std::move(cb); }
+    void set_closed_callback(NodeEventCallback cb)   { closed_callbacks_[0] = std::move(cb); }
+    void set_network_closed_callback(NodeEventCallback cb)   override { closed_callbacks_[1] = std::move(cb); }
+
+    const NodeStats& stats() const override { return stats_; }
+
+    NodeSnapshot node_snapshot(const std::string& name, double elapsed_s) const override {
+        uint64_t frames   = stats_.frames_processed.load(std::memory_order_relaxed);
+        double exec_ms    = stats_.ema_exec_us.load(std::memory_order_relaxed) / 1000.0;
+        double qwait_ms   = stats_.queue_wait_us.load(std::memory_order_relaxed) / 1000.0;
+        double total_ms   = exec_ms; // no blocked time for interrupt nodes
+        return {
+            name, frames, exec_ms,
+            stats_.max_exec_us.load(std::memory_order_relaxed) / 1000.0,
+            0.0,  // blocked_ms — not applicable
+            elapsed_s > 0 ? frames / elapsed_s : 0.0,
+            stats_.total_cpu_us.load(std::memory_order_relaxed) / 1000.0,
+            total_ms > 0 ? 100.0 : 0.0,
+            qwait_ms,
+        };
+    }
+
+    // ── Port access — by index ────────────────────────────────────────────────
+
+    template<std::size_t I>
+    OutputPort<InterruptNode, I> output() {
+        static_assert(I < output_count, "output index out of range");
+        return {*this};
+    }
+
+    template<fixed_string Name>
+    auto output() {
+        constexpr std::size_t idx = index_of<Name, OutNames...>();
+        static_assert(idx != npos, "unknown output port name");
+        return output<idx>();
+    }
+
+    template<std::size_t I>
+    void set_output_channel(Channel<std::tuple_element_t<I, return_tuple>>* ch) {
+        std::get<I>(output_channels_) = ch;
+    }
+
+    // ── Trigger ───────────────────────────────────────────────────────────────
+
+    // Returns a callable that fires this node when called.
+    // Pass it to a camera SDK, timer, or any external event source.
+    // Thread-safe; may be called from any thread.
+    std::function<void()> get_trigger() {
+        return [this] { trigger(); };
+    }
+
+private:
+    // Each trigger() increments pending_. When going 0→1 a task is submitted.
+    // Each fire_once() handles one pending event and decrements; if more remain
+    // (old value > 1) it resubmits itself. This guarantees every trigger produces
+    // exactly one execution even if triggers arrive faster than fire_once completes.
+    void trigger() {
+        if (stop_flag_.load(std::memory_order_relaxed)) return;
+        if (pending_.fetch_add(1, std::memory_order_acq_rel) == 0)
+            scheduler_->submit([this] { fire_once(); });
+    }
+
+    void fire_once() {
+        if (stop_flag_.load(std::memory_order_relaxed)) {
+            pending_.store(0, std::memory_order_release);
+            return;
+        }
+
+        auto t0 = clock_t::now();
+        int64_t now_us = std::chrono::duration_cast<std::chrono::microseconds>(
+            t0.time_since_epoch()).count();
+        stats_.exec_start_us.store(now_us, std::memory_order_relaxed);
+
+        bool fatal = false;
+        try {
+            auto t1 = clock_t::now();
+            stats_.record_queue_wait(duration_t(t1 - t0));
+            auto cpu0 = NodeStats::cpu_now();
+
+            if constexpr (std::is_void_v<return_raw>) {
+                Func();
+            } else {
+                auto result = Func();
+                push_outputs(normalise(std::move(result)),
+                             std::make_index_sequence<output_count>{});
+            }
+
+            auto cpu1 = NodeStats::cpu_now();
+            auto t2   = clock_t::now();
+            stats_.record_exec(duration_t(t2 - t1), duration_t::zero(), cpu0, cpu1);
+        } catch (const ChannelOverflowError&) {
+            fire_callbacks(event_callbacks_);
+        } catch (...) {
+            if (!error_handler_ || !error_handler_(name_, std::current_exception()))
+                fatal = true;
+        }
+
+        stats_.exec_start_us.store(0, std::memory_order_relaxed);
+
+        if (fatal) {
+            fire_callbacks(closed_callbacks_);
+            disable_outputs(std::make_index_sequence<output_count>{});
+            pending_.store(0, std::memory_order_release);
+            stop_flag_.store(true, std::memory_order_relaxed);
+            return;
+        }
+
+        // Decrement and resubmit only if more triggers are queued.
+        // fetch_sub returns old value; old > 1 means new > 0.
+        if (pending_.fetch_sub(1, std::memory_order_acq_rel) > 1)
+            scheduler_->submit([this] { fire_once(); });
+    }
+
+    template<typename R = return_raw>
+    static return_tuple normalise(R&& r) {
+        if constexpr (is_tuple_v<R>) return std::move(r);
+        else                         return std::make_tuple(std::move(r));
+    }
+
+    template<std::size_t... Is>
+    void push_outputs(return_tuple&& result, std::index_sequence<Is...>) {
+        (push_one<Is>(std::get<Is>(std::move(result))), ...);
+    }
+
+    template<std::size_t I>
+    void push_one(std::tuple_element_t<I, return_tuple>&& val) {
+        auto* ch = std::get<I>(output_channels_);
+        if (!ch) return;
+        try { ch->push(std::move(val)); }
+        catch (const ChannelOverflowError&) {
+            throw ChannelOverflowError(ch->capacity(),
+                "interrupt node '" + name_ + "' " + output_port_label<I>());
+        }
+    }
+
+    template<std::size_t I>
+    static std::string output_port_label() {
+        if constexpr (sizeof...(OutNames) > 0) {
+            constexpr std::array<std::string_view, sizeof...(OutNames)> names{OutNames.view()...};
+            return std::string("output['") + std::string(names[I]) + "']";
+        } else {
+            return "output[" + std::to_string(I) + "]";
+        }
+    }
+
+    template<typename Tup, std::size_t... Is>
+    static auto make_output_channel_tuple(std::index_sequence<Is...>)
+        -> std::tuple<Channel<std::tuple_element_t<Is, Tup>>*...>;
+
+    using output_channels_t = decltype(make_output_channel_tuple<return_tuple>(
+        std::make_index_sequence<output_count>{}));
+
+    template<std::size_t... Is>
+    void disable_outputs(std::index_sequence<Is...>) {
+        auto disable_one = [](auto* ch) { if (ch) ch->disable(); };
+        (disable_one(std::get<Is>(output_channels_)), ...);
+    }
+
+    static void fire_callbacks(const std::array<NodeEventCallback, 2>& cbs) {
+        const auto ts = std::chrono::steady_clock::now();
+        for (auto& cb : cbs) if (cb) cb(ts);
+    }
+
+    std::shared_ptr<IScheduler>              scheduler_;
+    std::string                              name_;
+    std::size_t                              fifo_capacity_;
+    output_channels_t                        output_channels_{};
+    std::atomic<bool>                        stop_flag_{true};
+    std::atomic<int>                         pending_{0};
+    NodeStats                                stats_;
+    NodeErrorHandler                         error_handler_;
+    std::chrono::milliseconds                max_exec_time_{0};
+    std::array<NodeEventCallback, 2>         event_callbacks_{};  // [0]=user [1]=network
+    std::array<NodeEventCallback, 2>         closed_callbacks_{};
+};
+
+// ── make_interrupt_node factory ───────────────────────────────────────────────
+
+template<auto Func, fixed_string Label = "", std::size_t UniqueTag = 0>
+auto make_interrupt_node(std::shared_ptr<IScheduler> sched, std::size_t fifo_capacity = 5) {
+    return InterruptNode<Func, out<>, Label, UniqueTag>(std::move(sched), fifo_capacity);
+}
+
+template<auto Func, fixed_string Label = "", std::size_t UniqueTag = 0,
+         fixed_string... OutNames>
+auto make_interrupt_node(std::shared_ptr<IScheduler> sched, out<OutNames...>,
+                         std::size_t fifo_capacity = 5) {
+    return InterruptNode<Func, out<OutNames...>, Label, UniqueTag>(
+        std::move(sched), fifo_capacity);
+}
+
+} // namespace kpn

include/kpn/kpn.hpp

@@ -5,8 +5,15 @@
 #include "traits.hpp"
 #include "channel.hpp"
 #include "port.hpp"
+#include "inode.hpp"
+#include "scheduler.hpp"
+#include "pool_node.hpp"
+#include "interrupt_node.hpp"
 #include "node.hpp"
 #include "fanout.hpp"
+#include "branch.hpp"
+#include "shared_resource.hpp"
 #include "static_network.hpp"
+#include "debug_hub.hpp"
 #include "main_thread_node.hpp"
 #include "network.hpp"

include/kpn/main_thread_node.hpp

@@ -2,7 +2,7 @@
 #include "channel.hpp"
 #include "diagnostics.hpp"
 #include "fixed_string.hpp"
-#include "node.hpp"       // INode
+#include "inode.hpp"
 #include "port.hpp"
 
 #include <atomic>

include/kpn/network.hpp

@@ -1,6 +1,6 @@
 #pragma once
 #include "diagnostics.hpp"
-#include "node.hpp"
+#include "inode.hpp"
 #include "port.hpp"
 
 #ifdef KPN_WEB_DEBUG
@@ -44,6 +44,9 @@ public:
     using DiagnosticsHandler =
         std::function<void(const std::vector<NodeSnapshot>&,
                            const std::vector<ChannelSnapshot>&)>;
+    using EventHandler =
+        std::function<void(std::string_view node_name, NodeEvent,
+                           std::chrono::steady_clock::time_point)>;
 
     // ── Builder API ───────────────────────────────────────────────────────────
 
@@ -111,6 +114,19 @@ public:
         for (auto& [name, _] : nodes_)
             if (color[name] == 0)
                 dfs(name, color);
+        if (event_handler_) {
+            for (auto& name : topo_) {
+                auto* node = nodes_.at(name);
+                node->set_network_overflow_callback(
+                    [this, n = name](auto ts) {
+                        event_handler_(n, NodeEvent::Overflow, ts);
+                    });
+                node->set_network_closed_callback(
+                    [this, n = name](auto ts) {
+                        event_handler_(n, NodeEvent::Closed, ts);
+                    });
+            }
+        }
         return *this;
     }
 
@@ -126,14 +142,17 @@ public:
             web_debug_port_,
             [this]() {
                 auto s = collect_snapshots();
-                return web_debug::to_json(s.nodes, s.channels, s.elapsed_s);
+                return web_debug::to_json(s.nodes, s.channels, {}, s.elapsed_s, s.pools);
             });
         web_server_->start();
         std::cerr << "[kpn] web debug UI: http://localhost:" << web_debug_port_ << "\n";
 #endif
     }
 
-    void stop() override {
+    void stop() override { halt(); }
+
+    // halt(): immediate stop — broadcasts disable to all channels and joins threads.
+    void halt() override {
 #ifdef KPN_WEB_DEBUG
         if (web_server_) web_server_->stop();
 #endif
@@ -142,6 +161,37 @@ public:
             nodes_.at(*it)->stop();
     }
 
+    // shutdown(): graceful drain in topological order.
+    // Stops source nodes first, polls until their output channels drain to zero,
+    // then stops the next layer, and so on.
+    void shutdown() override {
+#ifdef KPN_WEB_DEBUG
+        if (web_server_) web_server_->stop();
+#endif
+        stop_watchdog();
+
+        // Identify which nodes have no incoming edges (sources).
+        std::map<std::string, std::size_t> in_degree;
+        for (auto& [name, _] : nodes_) in_degree[name] = 0;
+        for (auto& [src, dsts] : adj_)
+            for (auto& dst : dsts) in_degree[dst]++;
+
+        // Walk topo order: stop each source layer, wait for its output channels
+        // to drain, then proceed to the next layer.
+        std::set<std::string> stopped;
+        for (auto& name : topo_) {
+            if (in_degree[name] == 0 || all_predecessors_stopped(name, stopped)) {
+                nodes_.at(name)->stop();
+                stopped.insert(name);
+                // Wait for output channels of this node to drain.
+                drain_output_channels(name);
+            }
+        }
+        // Stop any remaining nodes (sinks / nodes not yet stopped).
+        for (auto it = topo_.rbegin(); it != topo_.rend(); ++it)
+            if (!stopped.count(*it)) nodes_.at(*it)->stop();
+    }
+
     bool running() const override { return watchdog_.joinable(); }
     void set_name(std::string) override {}
 
@@ -160,8 +210,13 @@ public:
         watchdog_interval_ = interval;
     }
 
-    void set_error_handler(ErrorHandler h)       { error_handler_       = std::move(h); }
+    void set_error_handler(ErrorHandler h)             { error_handler_  = std::move(h); }
     void set_diagnostics_handler(DiagnosticsHandler h) { diag_handler_  = std::move(h); }
+    void set_event_handler(EventHandler h)             { event_handler_  = std::move(h); }
+
+    void register_pool(const std::string& name, IPoolProbe* probe) {
+        pool_probes_.emplace_back(name, probe);
+    }
 
 #ifdef KPN_WEB_DEBUG
     void set_web_debug_port(uint16_t port) { web_debug_port_ = port; }
@@ -171,7 +226,7 @@ public:
     // Can be called at any time; thread-safe (reads atomics with relaxed ordering).
     void print_diagnostics(std::ostream& os = std::cerr) const {
         auto s = collect_snapshots();
-        os << format_report(s.nodes, s.channels, s.elapsed_s);
+        os << format_report(s.nodes, s.channels, s.pools, s.elapsed_s);
     }
 
 private:
@@ -180,6 +235,7 @@ private:
     struct Snapshots {
         std::vector<NodeSnapshot>    nodes;
         std::vector<ChannelSnapshot> channels;
+        std::vector<PoolSnapshot>    pools;
         double                       elapsed_s;
     };
 
@@ -195,11 +251,16 @@ private:
         for (auto& probe : channel_probes_)
             channels.push_back(probe->snapshot());
 
-        return {std::move(nodes), std::move(channels), elapsed_s};
+        std::vector<PoolSnapshot> pools;
+        for (auto& [name, probe] : pool_probes_)
+            pools.push_back(probe->snapshot(name));
+
+        return {std::move(nodes), std::move(channels), std::move(pools), elapsed_s};
     }
 
     static std::string format_report(const std::vector<NodeSnapshot>& nodes,
                                      const std::vector<ChannelSnapshot>& channels,
+                                     const std::vector<PoolSnapshot>& pools = {},
                                      double elapsed_s = 0.0) {
         std::ostringstream os;
         os << std::fixed << std::setprecision(1);
@@ -258,6 +319,31 @@ private:
                << flag << "\n";
         }
 
+        // Pool table
+        if (!pools.empty()) {
+            os << "│\n│ Thread Pools:\n";
+            os << "│  " << std::left
+               << std::setw(16) << "name"
+               << std::setw(10) << "threads"
+               << std::setw(12) << "queued"
+               << std::setw(12) << "active"
+               << std::setw(14) << "in/s"
+               << std::setw(14) << "out/s"
+               << "\n│  " << std::string(78, '-') << "\n";
+            for (auto& p : pools) {
+                double in_rate  = elapsed_s > 0.0 ? p.tasks_submitted / elapsed_s : 0.0;
+                double out_rate = elapsed_s > 0.0 ? p.tasks_completed / elapsed_s : 0.0;
+                os << "│  " << std::left
+                   << std::setw(16) << p.name
+                   << std::setw(10) << p.thread_count
+                   << std::setw(12) << p.queue_depth
+                   << std::setw(12) << p.active_count
+                   << std::setw(14) << in_rate
+                   << std::setw(14) << out_rate
+                   << "\n";
+            }
+        }
+
         // Bottleneck hint: node with highest ema_exec_ms
         if (!nodes.empty()) {
             auto it = std::max_element(nodes.begin(), nodes.end(),
@@ -271,6 +357,31 @@ private:
         return os.str();
     }
 
+    // ── Shutdown helpers ──────────────────────────────────────────────────────
+
+    bool all_predecessors_stopped(const std::string& name,
+                                  const std::set<std::string>& stopped) const {
+        for (auto& [src, dsts] : adj_)
+            for (auto& dst : dsts)
+                if (dst == name && !stopped.count(src)) return false;
+        return true;
+    }
+
+    void drain_output_channels(const std::string& /*name*/) const {
+        // Poll all channel probes until none report non-zero fill.
+        // A short sleep prevents busy-spin; 1 ms is fine for drain purposes.
+        bool any_full = true;
+        while (any_full) {
+            any_full = false;
+            for (auto& probe : channel_probes_) {
+                auto snap = probe->snapshot();
+                if (snap.current_fill > 0) { any_full = true; break; }
+            }
+            if (any_full)
+                std::this_thread::sleep_for(std::chrono::milliseconds(1));
+        }
+    }
+
     // ── Cycle detection / topological sort ───────────────────────────────────
 
     void dfs(const std::string& name, std::map<std::string, int>& color) {
@@ -292,16 +403,33 @@ private:
                 if (tok.stop_requested()) break;
 
                 auto s = collect_snapshots();
+                check_hung_nodes();
 
                 if (diag_handler_) {
                     diag_handler_(s.nodes, s.channels);
                 } else {
-                    std::cerr << format_report(s.nodes, s.channels, s.elapsed_s);
+                    std::cerr << format_report(s.nodes, s.channels, s.pools, s.elapsed_s);
                 }
             }
         });
     }
 
+    void check_hung_nodes() const {
+        auto now_us = std::chrono::duration_cast<std::chrono::microseconds>(
+            clock_t::now().time_since_epoch()).count();
+        for (auto& [name, node] : nodes_) {
+            int64_t start = node->stats().exec_start_us.load(std::memory_order_relaxed);
+            if (start == 0) continue;
+            int64_t elapsed_ms = (now_us - start) / 1000;
+            // Warn if a node has been executing for > 5 s with no max_exec_time set,
+            // or if it exceeds its configured max. Threshold: 5000 ms default.
+            if (elapsed_ms > 5000) {
+                std::cerr << "[kpn] WARNING: node '" << name
+                          << "' has been executing for " << elapsed_ms << " ms\n";
+            }
+        }
+    }
+
     void stop_watchdog() {
         if (watchdog_.joinable())
             watchdog_.request_stop(), watchdog_.join();
@@ -316,8 +444,10 @@ private:
     std::map<std::string, std::string>                     exposed_outputs_;
     std::set<std::pair<std::string, std::size_t>>          connected_outputs_;
     std::vector<std::unique_ptr<IChannelProbe>>            channel_probes_;
+    std::vector<std::pair<std::string, IPoolProbe*>>       pool_probes_;
     ErrorHandler                                           error_handler_;
     DiagnosticsHandler                                     diag_handler_;
+    EventHandler                                           event_handler_;
     std::chrono::milliseconds                              watchdog_interval_{3000};
     std::jthread                                           watchdog_;
     clock_t::time_point                                    start_time_;

include/kpn/node.hpp

@@ -1,41 +1,29 @@
 #pragma once
-#include "channel.hpp"
-#include "diagnostics.hpp"
-#include "fixed_string.hpp"
-#include "port.hpp"
-#include "traits.hpp"
+#include "inode.hpp"
+#include "pool_node.hpp"   // PoolNode, PoolObjectNode
 
-#include <array>
-#include <atomic>
-#include <chrono>
-#include <cstddef>
-#include <iostream>
-#include <memory>
-#include <stdexcept>
-#include <thread>
-#include <tuple>
-#include <type_traits>
+// node.hpp — Node<> and ObjectNode<> as thin wrappers over PoolNode<>.
+//
+// Each Node owns a private single-thread ThreadPool so the API is unchanged:
+// node.start() / node.stop() are self-contained with no external scheduler.
+// Internally, all execution goes through PoolNode::fire_once() — the same code
+// path as explicitly pool-scheduled nodes.
+//
+// To share a thread pool across multiple nodes, use make_pool_node() directly.
 
 namespace kpn {
 
-// ── INode — type-erased interface for Network / watchdog ─────────────────────
+namespace detail {
 
-struct INode {
-    virtual ~INode()       = default;
-    virtual void start()   = 0;
-    virtual void stop()    = 0;
-    virtual bool running() const = 0;
-    virtual const NodeStats& stats() const = 0;
-    virtual NodeSnapshot node_snapshot(const std::string& name, double elapsed_s) const = 0;
-    virtual void set_name(std::string name) = 0;
+// Private base initialized before PoolNode so its pool can be passed to the
+// PoolNode constructor (C++ initialises bases left-to-right).
+struct NodePrivatePool {
+    std::shared_ptr<ThreadPool> pool{std::make_shared<ThreadPool>(1)};
 };
 
+} // namespace detail
+
 // ── Node ─────────────────────────────────────────────────────────────────────
-//
-// Template parameters:
-//   Func        — the wrapped function (auto NTTP, deduced as a function pointer)
-//   InputNames  — optional kpn::in<"a","b"> tag type (at most one)
-//   OutputNames — optional kpn::out<"x","y"> tag type (at most one)
 
 template<auto Func,
          typename InputTag  = in<>,
@@ -44,259 +32,25 @@ template<auto Func,
          std::size_t UniqueTag = 0>
 class Node;
 
-// Specialisation that unpacks the in<>/out<> tag packs
 template<auto Func, fixed_string... InNames, fixed_string... OutNames,
          fixed_string Label, std::size_t UniqueTag>
-class Node<Func, in<InNames...>, out<OutNames...>, Label, UniqueTag> : public INode {
+class Node<Func, in<InNames...>, out<OutNames...>, Label, UniqueTag>
+    : private detail::NodePrivatePool
+    , public  PoolNode<Func, in<InNames...>, out<OutNames...>, Label, UniqueTag> {
+    using Base = PoolNode<Func, in<InNames...>, out<OutNames...>, Label, UniqueTag>;
 public:
-    using F            = decltype(Func);
-    using args_tuple   = args_t<F>;
-    using return_raw   = return_t<F>;
-    using return_tuple = normalised_return_t<return_raw>;
-
-    // Identity accessors — used by StaticNetwork for diagnostics and type-level uniqueness
-    static constexpr std::string_view label()     { return Label.view(); }
-    static constexpr std::size_t      unique_tag  = UniqueTag;
-
-    static constexpr std::size_t input_count  = arity_v<F>;
-    static constexpr std::size_t output_count = std::tuple_size_v<return_tuple>;
-
-    static_assert(
-        sizeof...(InNames) == 0 || sizeof...(InNames) == input_count,
-        "make_node: number of input names must match function arity, or provide none"
-    );
-    static_assert(
-        sizeof...(OutNames) == 0 || sizeof...(OutNames) == output_count,
-        "make_node: number of output names must match return tuple size, or provide none"
-    );
-
     explicit Node(std::size_t fifo_capacity = 5)
-        : fifo_capacity_(fifo_capacity)
-    {
-        init_input_channels(std::make_index_sequence<input_count>{});
-    }
+        : detail::NodePrivatePool{}
+        , Base(pool, fifo_capacity)
+    {}
 
     ~Node() override { stop(); }
 
-    // ── INode ─────────────────────────────────────────────────────────────────
-
-    void start() override {
-        enable_inputs(std::make_index_sequence<input_count>{});
-        stop_flag_.store(false, std::memory_order_relaxed);
-        thread_ = std::jthread([this](std::stop_token) { run_loop(); });
-    }
-
-    void stop() override {
-        stop_flag_.store(true, std::memory_order_relaxed);
-        // Disable all input channels: drops queued items and unblocks waiting pop()
-        disable_inputs(std::make_index_sequence<input_count>{});
-        if (thread_.joinable()) thread_.request_stop(), thread_.join();
-    }
-
-    bool running() const override {
-        return thread_.joinable() && !stop_flag_.load(std::memory_order_relaxed);
-    }
-
-    void set_name(std::string name) override { name_ = std::move(name); }
-
-    const NodeStats& stats() const override { return stats_; }
-
-    NodeSnapshot node_snapshot(const std::string& name, double elapsed_s) const override {
-        uint64_t frames   = stats_.frames_processed.load(std::memory_order_relaxed);
-        double exec_ms    = stats_.ema_exec_us.load(std::memory_order_relaxed) / 1000.0;
-        double blocked_ms = stats_.total_blocked_us.load(std::memory_order_relaxed) / 1000.0;
-        double total_ms   = exec_ms + blocked_ms;
-        return {
-            name,
-            frames,
-            exec_ms,
-            stats_.max_exec_us.load(std::memory_order_relaxed) / 1000.0,
-            blocked_ms,
-            elapsed_s > 0 ? frames / elapsed_s : 0.0,
-            stats_.total_cpu_us.load(std::memory_order_relaxed) / 1000.0,
-            total_ms > 0 ? 100.0 * exec_ms / total_ms : 0.0,
-        };
-    }
-
-    // ── Port access — by index ────────────────────────────────────────────────
-
-    template<std::size_t I>
-    InputPort<Node, I> input() {
-        static_assert(I < input_count, "input index out of range");
-        return {*this};
-    }
-
-    template<std::size_t I>
-    OutputPort<Node, I> output() {
-        static_assert(I < output_count, "output index out of range");
-        return {*this};
-    }
-
-    // ── Port access — by name ─────────────────────────────────────────────────
-
-    template<fixed_string Name>
-    auto input() {
-        constexpr std::size_t idx = index_of<Name, InNames...>();
-        static_assert(idx != npos, "unknown input port name");
-        return input<idx>();
-    }
-
-    template<fixed_string Name>
-    auto output() {
-        constexpr std::size_t idx = index_of<Name, OutNames...>();
-        static_assert(idx != npos, "unknown output port name");
-        return output<idx>();
-    }
-
-    // ── Internal channel accessors (used by Network at connect time) ──────────
-
-    template<std::size_t I>
-    Channel<std::tuple_element_t<I, args_tuple>>& input_channel() {
-        return *std::get<I>(input_channels_);
-    }
-
-    // Replace the owned input channel with an externally provided one.
-    // Used by VariantNodeWrapper to share a Channel<T> with a VariantChannel adapter.
-    template<std::size_t I>
-    void set_input_channel(
-        std::shared_ptr<Channel<std::tuple_element_t<I, args_tuple>>> ch) {
-        std::get<I>(input_channels_) = std::move(ch);
-    }
-
-    template<std::size_t I>
-    void set_output_channel(
-        Channel<std::tuple_element_t<I, return_tuple>>* ch) {
-        std::get<I>(output_channels_) = ch;
-    }
-
-private:
-    // ── Channel storage ───────────────────────────────────────────────────────
-
-    // Input channels — shared ownership so VariantChannel adapters can share them
-    template<std::size_t... Is>
-    void init_input_channels(std::index_sequence<Is...>) {
-        ((std::get<Is>(input_channels_) =
-              std::make_shared<Channel<std::tuple_element_t<Is, args_tuple>>>(fifo_capacity_)),
-         ...);
-    }
-
-    template<std::size_t... Is>
-    void enable_inputs(std::index_sequence<Is...>) {
-        (std::get<Is>(input_channels_)->enable(), ...);
-    }
-
-    template<std::size_t... Is>
-    void disable_inputs(std::index_sequence<Is...>) {
-        (std::get<Is>(input_channels_)->disable(), ...);
-    }
-
-    template<typename Tup, std::size_t... Is>
-    static auto make_input_channel_tuple(std::index_sequence<Is...>)
-        -> std::tuple<std::shared_ptr<Channel<std::tuple_element_t<Is, Tup>>>...>;
-
-    using input_channels_t  = decltype(make_input_channel_tuple<args_tuple>(
-        std::make_index_sequence<input_count>{}));
-
-    // Output channels — non-owning pointers, set at connect time
-    template<typename Tup, std::size_t... Is>
-    static auto make_output_channel_tuple(std::index_sequence<Is...>)
-        -> std::tuple<Channel<std::tuple_element_t<Is, Tup>>*...>;
-
-    using output_channels_t = decltype(make_output_channel_tuple<return_tuple>(
-        std::make_index_sequence<output_count>{}));
-
-    // ── run_loop ──────────────────────────────────────────────────────────────
-
-    void run_loop() {
-        while (!stop_flag_.load(std::memory_order_relaxed)) {
-            try {
-                auto t0 = clock_t::now();
-                auto args = pop_inputs(std::make_index_sequence<input_count>{});
-                auto t1 = clock_t::now();
-                auto cpu0 = NodeStats::cpu_now();
-
-                if constexpr (std::is_void_v<return_raw>) {
-                    std::apply(Func, args);
-                } else {
-                    auto result = std::apply(Func, args);
-                    push_outputs(normalise(std::move(result)),
-                                 std::make_index_sequence<output_count>{});
-                }
-
-                auto cpu1 = NodeStats::cpu_now();
-                auto t2 = clock_t::now();
-                stats_.record_exec(duration_t(t2 - t1), duration_t(t1 - t0), cpu0, cpu1);
-            } catch (const ChannelClosedError&) {
-                break;
-            } catch (const ChannelOverflowError& e) {
-                std::cerr << "[kpn] overflow: " << e.what() << "\n";
-            }
-        }
-    }
-
-    // Pop all inputs into a tuple of argument values
-    template<std::size_t... Is>
-    args_tuple pop_inputs(std::index_sequence<Is...>) {
-        return {std::get<Is>(input_channels_)->pop()...};
-    }
-
-    // Normalise return value to tuple (handles void and single-value returns)
-    template<typename R = return_raw>
-    static return_tuple normalise(R&& r) {
-        if constexpr (is_tuple_v<R>)
-            return std::move(r);
-        else
-            return std::make_tuple(std::move(r));
-    }
-
-    static return_tuple normalise_void() { return {}; }
-
-    // Push each output element to its connected channel (if connected)
-    template<std::size_t... Is>
-    void push_outputs(return_tuple&& result, std::index_sequence<Is...>) {
-        (push_one<Is>(std::get<Is>(std::move(result))), ...);
-    }
-
-    template<std::size_t I>
-    void push_one(std::tuple_element_t<I, return_tuple>&& val) {
-        auto* ch = std::get<I>(output_channels_);
-        if (!ch) return;
-        try {
-            ch->push(std::move(val));
-        } catch (const ChannelOverflowError&) {
-            throw ChannelOverflowError(ch->capacity(), "node '" + name_ + "' " + output_port_label<I>());
-        }
-    }
-
-    template<std::size_t I>
-    static std::string output_port_label() {
-        if constexpr (sizeof...(OutNames) > 0) {
-            constexpr std::array<std::string_view, sizeof...(OutNames)> names{OutNames.view()...};
-            return std::string("output['") + std::string(names[I]) + "']";
-        } else {
-            return "output[" + std::to_string(I) + "]";
-        }
-    }
-
-    // ── State ─────────────────────────────────────────────────────────────────
-
-    std::string         name_;
-    std::size_t         fifo_capacity_;
-    input_channels_t    input_channels_;
-    output_channels_t   output_channels_{};
-    std::atomic<bool>   stop_flag_{false};
-    std::jthread        thread_;
-    NodeStats           stats_;
+    void start() override { pool->start(); Base::start(); }
+    void stop()  override { Base::stop();  pool->stop();  }
 };
 
-// ── ObjectNode — wraps a callable object (functor / class with operator()) ────
-//
-// Use this when the node needs state initialised in a constructor.
-// The object must outlive the ObjectNode (stored by reference).
-//
-// Usage:
-//   MyFunctor obj(...);
-//   auto node = make_node(obj, in<"x">{}, out<"y">{}, capacity);
+// ── ObjectNode ────────────────────────────────────────────────────────────────
 
 template<typename Obj,
          typename InputTag  = in<>,
@@ -307,222 +61,20 @@ class ObjectNode;
 
 template<typename Obj, fixed_string... InNames, fixed_string... OutNames,
          fixed_string Label, std::size_t UniqueTag>
-class ObjectNode<Obj, in<InNames...>, out<OutNames...>, Label, UniqueTag> : public INode {
+class ObjectNode<Obj, in<InNames...>, out<OutNames...>, Label, UniqueTag>
+    : private detail::NodePrivatePool
+    , public  PoolObjectNode<Obj, in<InNames...>, out<OutNames...>, Label, UniqueTag> {
+    using Base = PoolObjectNode<Obj, in<InNames...>, out<OutNames...>, Label, UniqueTag>;
 public:
-    using F            = decltype(&Obj::operator());
-    using args_tuple   = args_t<F>;
-    using return_raw   = return_t<F>;
-    using return_tuple = normalised_return_t<return_raw>;
-
-    static constexpr std::string_view label()    { return Label.view(); }
-    static constexpr std::size_t      unique_tag = UniqueTag;
-
-    static constexpr std::size_t input_count  = arity_v<F>;
-    static constexpr std::size_t output_count = std::tuple_size_v<return_tuple>;
-
-    static_assert(
-        sizeof...(InNames) == 0 || sizeof...(InNames) == input_count,
-        "make_node: number of input names must match operator() arity, or provide none"
-    );
-    static_assert(
-        sizeof...(OutNames) == 0 || sizeof...(OutNames) == output_count,
-        "make_node: number of output names must match return tuple size, or provide none"
-    );
-
     explicit ObjectNode(Obj& obj, std::size_t fifo_capacity = 5)
-        : obj_(obj), fifo_capacity_(fifo_capacity)
-    {
-        init_input_channels(std::make_index_sequence<input_count>{});
-    }
+        : detail::NodePrivatePool{}
+        , Base(obj, pool, fifo_capacity)
+    {}
 
     ~ObjectNode() override { stop(); }
 
-    // ── INode ─────────────────────────────────────────────────────────────────
-
-    void start() override {
-        enable_inputs(std::make_index_sequence<input_count>{});
-        stop_flag_.store(false, std::memory_order_relaxed);
-        thread_ = std::jthread([this](std::stop_token) { run_loop(); });
-    }
-
-    void stop() override {
-        stop_flag_.store(true, std::memory_order_relaxed);
-        disable_inputs(std::make_index_sequence<input_count>{});
-        if (thread_.joinable()) thread_.request_stop(), thread_.join();
-    }
-
-    bool running() const override {
-        return thread_.joinable() && !stop_flag_.load(std::memory_order_relaxed);
-    }
-
-    void set_name(std::string name) override { name_ = std::move(name); }
-
-    const NodeStats& stats() const override { return stats_; }
-
-    NodeSnapshot node_snapshot(const std::string& name, double elapsed_s) const override {
-        uint64_t frames   = stats_.frames_processed.load(std::memory_order_relaxed);
-        double exec_ms    = stats_.ema_exec_us.load(std::memory_order_relaxed) / 1000.0;
-        double blocked_ms = stats_.total_blocked_us.load(std::memory_order_relaxed) / 1000.0;
-        double total_ms   = exec_ms + blocked_ms;
-        return {
-            name, frames,
-            exec_ms,
-            stats_.max_exec_us.load(std::memory_order_relaxed) / 1000.0,
-            blocked_ms,
-            elapsed_s > 0 ? frames / elapsed_s : 0.0,
-            stats_.total_cpu_us.load(std::memory_order_relaxed) / 1000.0,
-            total_ms > 0 ? 100.0 * exec_ms / total_ms : 0.0,
-        };
-    }
-
-    // ── Port access ───────────────────────────────────────────────────────────
-
-    template<std::size_t I>
-    InputPort<ObjectNode, I> input() {
-        static_assert(I < input_count, "input index out of range");
-        return {*this};
-    }
-
-    template<std::size_t I>
-    OutputPort<ObjectNode, I> output() {
-        static_assert(I < output_count, "output index out of range");
-        return {*this};
-    }
-
-    template<fixed_string Name>
-    auto input() {
-        constexpr std::size_t idx = index_of<Name, InNames...>();
-        static_assert(idx != npos, "unknown input port name");
-        return input<idx>();
-    }
-
-    template<fixed_string Name>
-    auto output() {
-        constexpr std::size_t idx = index_of<Name, OutNames...>();
-        static_assert(idx != npos, "unknown output port name");
-        return output<idx>();
-    }
-
-    template<std::size_t I>
-    Channel<std::tuple_element_t<I, args_tuple>>& input_channel() {
-        return *std::get<I>(input_channels_);
-    }
-
-    template<std::size_t I>
-    void set_input_channel(
-        std::shared_ptr<Channel<std::tuple_element_t<I, args_tuple>>> ch) {
-        std::get<I>(input_channels_) = std::move(ch);
-    }
-
-    template<std::size_t I>
-    void set_output_channel(Channel<std::tuple_element_t<I, return_tuple>>* ch) {
-        std::get<I>(output_channels_) = ch;
-    }
-
-private:
-    template<std::size_t... Is>
-    void init_input_channels(std::index_sequence<Is...>) {
-        ((std::get<Is>(input_channels_) =
-              std::make_shared<Channel<std::tuple_element_t<Is, args_tuple>>>(fifo_capacity_)),
-         ...);
-    }
-
-    template<std::size_t... Is>
-    void enable_inputs(std::index_sequence<Is...>) {
-        (std::get<Is>(input_channels_)->enable(), ...);
-    }
-
-    template<std::size_t... Is>
-    void disable_inputs(std::index_sequence<Is...>) {
-        (std::get<Is>(input_channels_)->disable(), ...);
-    }
-
-    template<typename Tup, std::size_t... Is>
-    static auto make_input_channel_tuple(std::index_sequence<Is...>)
-        -> std::tuple<std::shared_ptr<Channel<std::tuple_element_t<Is, Tup>>>...>;
-
-    using input_channels_t = decltype(make_input_channel_tuple<args_tuple>(
-        std::make_index_sequence<input_count>{}));
-
-    template<typename Tup, std::size_t... Is>
-    static auto make_output_channel_tuple(std::index_sequence<Is...>)
-        -> std::tuple<Channel<std::tuple_element_t<Is, Tup>>*...>;
-
-    using output_channels_t = decltype(make_output_channel_tuple<return_tuple>(
-        std::make_index_sequence<output_count>{}));
-
-    void run_loop() {
-        while (!stop_flag_.load(std::memory_order_relaxed)) {
-            try {
-                auto t0   = clock_t::now();
-                auto args = pop_inputs(std::make_index_sequence<input_count>{});
-                auto t1   = clock_t::now();
-                auto cpu0 = NodeStats::cpu_now();
-
-                if constexpr (std::is_void_v<return_raw>) {
-                    std::apply([this](auto&&... a) { obj_(std::forward<decltype(a)>(a)...); }, args);
-                } else {
-                    auto result = std::apply([this](auto&&... a) { return obj_(std::forward<decltype(a)>(a)...); }, args);
-                    push_outputs(normalise(std::move(result)),
-                                 std::make_index_sequence<output_count>{});
-                }
-
-                auto cpu1 = NodeStats::cpu_now();
-                auto t2   = clock_t::now();
-                stats_.record_exec(duration_t(t2 - t1), duration_t(t1 - t0), cpu0, cpu1);
-            } catch (const ChannelClosedError&) {
-                break;
-            } catch (const ChannelOverflowError& e) {
-                std::cerr << "[kpn] overflow: " << e.what() << "\n";
-            }
-        }
-    }
-
-    template<std::size_t... Is>
-    args_tuple pop_inputs(std::index_sequence<Is...>) {
-        return {std::get<Is>(input_channels_)->pop()...};
-    }
-
-    template<typename R = return_raw>
-    static return_tuple normalise(R&& r) {
-        if constexpr (is_tuple_v<R>) return std::move(r);
-        else                         return std::make_tuple(std::move(r));
-    }
-
-    template<std::size_t... Is>
-    void push_outputs(return_tuple&& result, std::index_sequence<Is...>) {
-        (push_one<Is>(std::get<Is>(std::move(result))), ...);
-    }
-
-    template<std::size_t I>
-    void push_one(std::tuple_element_t<I, return_tuple>&& val) {
-        auto* ch = std::get<I>(output_channels_);
-        if (!ch) return;
-        try {
-            ch->push(std::move(val));
-        } catch (const ChannelOverflowError&) {
-            throw ChannelOverflowError(ch->capacity(), "node '" + name_ + "' " + output_port_label<I>());
-        }
-    }
-
-    template<std::size_t I>
-    static std::string output_port_label() {
-        if constexpr (sizeof...(OutNames) > 0) {
-            constexpr std::array<std::string_view, sizeof...(OutNames)> names{OutNames.view()...};
-            return std::string("output['") + std::string(names[I]) + "']";
-        } else {
-            return "output[" + std::to_string(I) + "]";
-        }
-    }
-
-    Obj&                obj_;
-    std::string         name_;
-    std::size_t         fifo_capacity_;
-    input_channels_t    input_channels_;
-    output_channels_t   output_channels_{};
-    std::atomic<bool>   stop_flag_{false};
-    std::jthread        thread_;
-    NodeStats           stats_;
+    void start() override { pool->start(); Base::start(); }
+    void stop()  override { Base::stop();  pool->stop();  }
 };
 
 // ── make_node overloads for callable objects ──────────────────────────────────
@@ -548,35 +100,24 @@ auto make_node(Obj& obj, in<InNames...>, out<OutNames...>, std::size_t fifo_capa
 }
 
 // ── make_node factory (NTTP) ──────────────────────────────────────────────────
-//
-// Usage:
-//   make_node<func>(capacity)
-//   make_node<func, "label">(capacity)
-//   make_node<func, "label", 1>(capacity)          // UniqueTag=1
-//   make_node<func>(in<"a","b">{}, capacity)
-//   make_node<func, "label">(in<"a","b">{}, out<"x">{}, capacity)
 
-// No port names
 template<auto Func, fixed_string Label = "", std::size_t UniqueTag = 0>
 auto make_node(std::size_t fifo_capacity = 5) {
     return Node<Func, in<>, out<>, Label, UniqueTag>(fifo_capacity);
 }
 
-// in<> only
 template<auto Func, fixed_string Label = "", std::size_t UniqueTag = 0,
          fixed_string... InNames>
 auto make_node(in<InNames...>, std::size_t fifo_capacity = 5) {
     return Node<Func, in<InNames...>, out<>, Label, UniqueTag>(fifo_capacity);
 }
 
-// out<> only
 template<auto Func, fixed_string Label = "", std::size_t UniqueTag = 0,
          fixed_string... OutNames>
 auto make_node(out<OutNames...>, std::size_t fifo_capacity = 5) {
     return Node<Func, in<>, out<OutNames...>, Label, UniqueTag>(fifo_capacity);
 }
 
-// in<> and out<>
 template<auto Func, fixed_string Label = "", std::size_t UniqueTag = 0,
          fixed_string... InNames, fixed_string... OutNames>
 auto make_node(in<InNames...>, out<OutNames...>, std::size_t fifo_capacity = 5) {

include/kpn/pool_node.hpp

@@ -0,0 +1,743 @@
+#pragma once
+#include "channel.hpp"
+#include "diagnostics.hpp"
+#include "fixed_string.hpp"
+#include "inode.hpp"
+#include "port.hpp"
+#include "scheduler.hpp"
+#include "traits.hpp"
+
+#include <array>
+#include <atomic>
+#include <chrono>
+#include <cstddef>
+#include <functional>
+#include <iostream>
+#include <memory>
+#include <optional>
+#include <stdexcept>
+#include <thread>
+#include <tuple>
+#include <type_traits>
+
+namespace kpn {
+
+// ── PoolNode ──────────────────────────────────────────────────────────────────
+//
+// Reactive alternative to Node<>. Instead of owning a blocked thread, the node
+// is submitted to a shared IScheduler whenever all its input channels become
+// non-empty. A single fire_once() call pops all inputs, executes the function,
+// and pushes outputs. At most one fire_once() runs at a time (queued_ flag).
+//
+// Source nodes (input_count == 0) submit themselves immediately on start() and
+// resubmit after each fire_once().
+//
+// Multiple PoolNodes can share one ThreadPool for resource-bounded execution,
+// or each can have a dedicated single-thread pool for serialisation.
+
+template<auto Func,
+         typename InputTag  = in<>,
+         typename OutputTag = out<>,
+         fixed_string Label    = "",
+         std::size_t UniqueTag = 0>
+class PoolNode;
+
+template<auto Func, fixed_string... InNames, fixed_string... OutNames,
+         fixed_string Label, std::size_t UniqueTag>
+class PoolNode<Func, in<InNames...>, out<OutNames...>, Label, UniqueTag> : public INode {
+public:
+    using F            = decltype(Func);
+    using args_tuple   = args_t<F>;
+    using return_raw   = return_t<F>;
+    using return_tuple = normalised_return_t<return_raw>;
+
+    static constexpr std::string_view label()     { return Label.view(); }
+    static constexpr std::size_t      unique_tag  = UniqueTag;
+
+    static constexpr std::size_t input_count  = arity_v<F>;
+    static constexpr std::size_t output_count = std::tuple_size_v<return_tuple>;
+
+    static_assert(
+        sizeof...(InNames) == 0 || sizeof...(InNames) == input_count,
+        "make_pool_node: number of input names must match function arity, or provide none"
+    );
+    static_assert(
+        sizeof...(OutNames) == 0 || sizeof...(OutNames) == output_count,
+        "make_pool_node: number of output names must match return tuple size, or provide none"
+    );
+
+    explicit PoolNode(std::shared_ptr<IScheduler> sched, std::size_t fifo_capacity = 5)
+        : scheduler_(std::move(sched)), fifo_capacity_(fifo_capacity)
+    {
+        init_input_channels(std::make_index_sequence<input_count>{});
+    }
+
+    ~PoolNode() override { stop(); }
+
+    // ── INode ─────────────────────────────────────────────────────────────────
+
+    void start() override {
+        enable_inputs(std::make_index_sequence<input_count>{});
+        stop_flag_.store(false, std::memory_order_relaxed);
+        queued_.store(false, std::memory_order_relaxed);
+        register_callbacks(std::make_index_sequence<input_count>{});
+        if constexpr (input_count == 0)
+            try_submit(0.5f);
+    }
+
+    void stop() override {
+        stop_flag_.store(true, std::memory_order_seq_cst);
+        disable_inputs(std::make_index_sequence<input_count>{});
+        // fire_once() observes stop_flag_ and will not resubmit.
+        // We do not wait for an in-flight fire_once() to complete here;
+        // callers that need that guarantee should call scheduler_->drain() first.
+    }
+
+    bool running() const override {
+        return !stop_flag_.load(std::memory_order_relaxed);
+    }
+
+    void set_name(std::string name) override { name_ = std::move(name); }
+    void set_error_handler(NodeErrorHandler h) { error_handler_ = std::move(h); }
+    void set_max_exec_time(std::chrono::milliseconds t) { max_exec_time_ = t; }
+
+    void set_overflow_callback(NodeEventCallback cb) { event_callbacks_[0] = std::move(cb); }
+    void set_network_overflow_callback(NodeEventCallback cb) override { event_callbacks_[1] = std::move(cb); }
+    void set_closed_callback(NodeEventCallback cb)   { closed_callbacks_[0] = std::move(cb); }
+    void set_network_closed_callback(NodeEventCallback cb)   override { closed_callbacks_[1] = std::move(cb); }
+
+    const NodeStats& stats() const override { return stats_; }
+
+    NodeSnapshot node_snapshot(const std::string& name, double elapsed_s) const override {
+        uint64_t frames    = stats_.frames_processed.load(std::memory_order_relaxed);
+        double exec_ms     = stats_.ema_exec_us.load(std::memory_order_relaxed) / 1000.0;
+        double blocked_ms  = stats_.total_blocked_us.load(std::memory_order_relaxed) / 1000.0;
+        double qwait_ms    = stats_.queue_wait_us.load(std::memory_order_relaxed) / 1000.0;
+        double total_ms    = exec_ms + blocked_ms;
+        return {
+            name, frames, exec_ms,
+            stats_.max_exec_us.load(std::memory_order_relaxed) / 1000.0,
+            blocked_ms,
+            elapsed_s > 0 ? frames / elapsed_s : 0.0,
+            stats_.total_cpu_us.load(std::memory_order_relaxed) / 1000.0,
+            total_ms > 0 ? 100.0 * exec_ms / total_ms : 0.0,
+            qwait_ms,
+        };
+    }
+
+    // ── Port access — by index ────────────────────────────────────────────────
+
+    template<std::size_t I>
+    InputPort<PoolNode, I> input() {
+        static_assert(I < input_count, "input index out of range");
+        return {*this};
+    }
+
+    template<std::size_t I>
+    OutputPort<PoolNode, I> output() {
+        static_assert(I < output_count, "output index out of range");
+        return {*this};
+    }
+
+    // ── Port access — by name ─────────────────────────────────────────────────
+
+    template<fixed_string Name>
+    auto input() {
+        constexpr std::size_t idx = index_of<Name, InNames...>();
+        static_assert(idx != npos, "unknown input port name");
+        return input<idx>();
+    }
+
+    template<fixed_string Name>
+    auto output() {
+        constexpr std::size_t idx = index_of<Name, OutNames...>();
+        static_assert(idx != npos, "unknown output port name");
+        return output<idx>();
+    }
+
+    // ── Internal channel accessors ────────────────────────────────────────────
+
+    template<std::size_t I>
+    Channel<std::tuple_element_t<I, args_tuple>>& input_channel() {
+        return *std::get<I>(input_channels_);
+    }
+
+    template<std::size_t I>
+    void set_input_channel(
+        std::shared_ptr<Channel<std::tuple_element_t<I, args_tuple>>> ch) {
+        std::get<I>(input_channels_) = std::move(ch);
+    }
+
+    template<std::size_t I>
+    void set_output_channel(
+        Channel<std::tuple_element_t<I, return_tuple>>* ch) {
+        std::get<I>(output_channels_) = ch;
+    }
+
+private:
+    // ── Channel storage ───────────────────────────────────────────────────────
+
+    template<std::size_t... Is>
+    void init_input_channels(std::index_sequence<Is...>) {
+        ((std::get<Is>(input_channels_) =
+              std::make_shared<Channel<std::tuple_element_t<Is, args_tuple>>>(fifo_capacity_)),
+         ...);
+    }
+
+    template<std::size_t... Is>
+    void enable_inputs(std::index_sequence<Is...>) {
+        (std::get<Is>(input_channels_)->enable(), ...);
+    }
+
+    template<std::size_t... Is>
+    void disable_inputs(std::index_sequence<Is...>) {
+        (std::get<Is>(input_channels_)->disable(), ...);
+    }
+
+    template<std::size_t... Is>
+    void disable_outputs(std::index_sequence<Is...>) {
+        auto disable_one = [](auto* ch) { if (ch) ch->disable(); };
+        (disable_one(std::get<Is>(output_channels_)), ...);
+    }
+
+    template<std::size_t... Is>
+    void register_callbacks(std::index_sequence<Is...>) {
+        (std::get<Is>(input_channels_)->set_push_callback(
+            [this] { on_input_ready(); }), ...);
+    }
+
+    static void fire_callbacks(const std::array<NodeEventCallback, 2>& cbs) {
+        const auto ts = std::chrono::steady_clock::now();
+        for (auto& cb : cbs) if (cb) cb(ts);
+    }
+
+    void self_stop() {
+        disable_inputs(std::make_index_sequence<input_count>{});
+        disable_outputs(std::make_index_sequence<output_count>{});
+        stats_.exec_start_us.store(0, std::memory_order_relaxed);
+        queued_.store(false, std::memory_order_release);
+        stop_flag_.store(true, std::memory_order_relaxed);
+    }
+
+    template<typename Tup, std::size_t... Is>
+    static auto make_input_channel_tuple(std::index_sequence<Is...>)
+        -> std::tuple<std::shared_ptr<Channel<std::tuple_element_t<Is, Tup>>>...>;
+
+    using input_channels_t = decltype(make_input_channel_tuple<args_tuple>(
+        std::make_index_sequence<input_count>{}));
+
+    template<typename Tup, std::size_t... Is>
+    static auto make_output_channel_tuple(std::index_sequence<Is...>)
+        -> std::tuple<Channel<std::tuple_element_t<Is, Tup>>*...>;
+
+    using output_channels_t = decltype(make_output_channel_tuple<return_tuple>(
+        std::make_index_sequence<output_count>{}));
+
+    // ── Scheduling ────────────────────────────────────────────────────────────
+
+    // Called by channel push_callbacks (on the producer's thread).
+    void on_input_ready() {
+        if (stop_flag_.load(std::memory_order_relaxed)) return;
+        std::size_t ready = count_ready(std::make_index_sequence<input_count>{});
+        if (ready == input_count)
+            try_submit(compute_priority());
+    }
+
+    template<std::size_t... Is>
+    std::size_t count_ready(std::index_sequence<Is...>) {
+        return ((std::get<Is>(input_channels_)->approx_size() > 0 ? 1u : 0u) + ...);
+    }
+
+    float compute_priority() {
+        if constexpr (input_count == 0) return 0.5f;
+        float sum = 0.0f;
+        sum_fill(sum, std::make_index_sequence<input_count>{});
+        return sum / static_cast<float>(input_count);
+    }
+
+    template<std::size_t... Is>
+    void sum_fill(float& sum, std::index_sequence<Is...>) {
+        ((sum += std::get<Is>(input_channels_)->capacity() > 0
+            ? float(std::get<Is>(input_channels_)->approx_size())
+              / float(std::get<Is>(input_channels_)->capacity())
+            : 0.5f), ...);
+    }
+
+    void try_submit(float priority) {
+        bool expected = false;
+        if (queued_.compare_exchange_strong(expected, true, std::memory_order_acq_rel))
+            scheduler_->submit([this] { fire_once(); }, priority);
+    }
+
+    // ── Execution ─────────────────────────────────────────────────────────────
+
+    void fire_once() {
+        if (stop_flag_.load(std::memory_order_relaxed)) {
+            queued_.store(false, std::memory_order_release);
+            return;
+        }
+
+        // Record queue wait time (submission → now) and mark as executing
+        auto t0 = clock_t::now();
+        int64_t now_us = std::chrono::duration_cast<std::chrono::microseconds>(
+            t0.time_since_epoch()).count();
+        stats_.exec_start_us.store(now_us, std::memory_order_relaxed);
+
+        try {
+            auto args = pop_inputs(std::make_index_sequence<input_count>{});
+            auto t1 = clock_t::now();
+            stats_.record_queue_wait(duration_t(t1 - t0));
+            auto cpu0 = NodeStats::cpu_now();
+
+            if constexpr (std::is_void_v<return_raw>) {
+                std::apply(Func, args);
+            } else {
+                auto result = std::apply(Func, args);
+                push_outputs(normalise(std::move(result)),
+                             std::make_index_sequence<output_count>{});
+            }
+
+            auto cpu1 = NodeStats::cpu_now();
+            auto t2   = clock_t::now();
+            // blocked_time = 0 for pool nodes (we don't block waiting for inputs)
+            stats_.record_exec(duration_t(t2 - t1), duration_t::zero(), cpu0, cpu1);
+        } catch (const ChannelClosedError&) {
+            fire_callbacks(closed_callbacks_);
+            self_stop();
+            return;
+        } catch (const ChannelOverflowError&) {
+            fire_callbacks(event_callbacks_);
+        } catch (...) {
+            if (error_handler_ && error_handler_(name_, std::current_exception())) {
+                // continue — fall through to resubmit check
+            } else {
+                fire_callbacks(closed_callbacks_);
+                self_stop();
+                return;
+            }
+        }
+
+        stats_.exec_start_us.store(0, std::memory_order_relaxed);
+        queued_.store(false, std::memory_order_release);
+
+        if (stop_flag_.load(std::memory_order_relaxed)) return;
+
+        // Source nodes always resubmit; others resubmit only if inputs are ready.
+        if constexpr (input_count == 0) {
+            try_submit(0.5f);
+        } else {
+            on_input_ready();
+        }
+    }
+
+    // Pop all inputs — safe because we're the sole consumer and fire_once
+    // is guarded by queued_ (only one fire_once runs at a time).
+    template<std::size_t... Is>
+    args_tuple pop_inputs(std::index_sequence<Is...>) {
+        return {pop_one<Is>()...};
+    }
+
+    template<std::size_t I>
+    std::tuple_element_t<I, args_tuple> pop_one() {
+        auto& ch = *std::get<I>(input_channels_);
+        std::tuple_element_t<I, args_tuple> val;
+        if (!ch.try_pop_now(val))
+            throw ChannelClosedError{};
+        return val;
+    }
+
+    template<typename R = return_raw>
+    static return_tuple normalise(R&& r) {
+        if constexpr (is_tuple_v<R>) return std::move(r);
+        else                         return std::make_tuple(std::move(r));
+    }
+
+    template<std::size_t... Is>
+    void push_outputs(return_tuple&& result, std::index_sequence<Is...>) {
+        (push_one_out<Is>(std::get<Is>(std::move(result))), ...);
+    }
+
+    template<std::size_t I>
+    void push_one_out(std::tuple_element_t<I, return_tuple>&& val) {
+        auto* ch = std::get<I>(output_channels_);
+        if (!ch) return;
+        try {
+            ch->push(std::move(val));
+        } catch (const ChannelOverflowError&) {
+            throw ChannelOverflowError(ch->capacity(),
+                "pool node '" + name_ + "' " + output_port_label<I>());
+        }
+    }
+
+    template<std::size_t I>
+    static std::string output_port_label() {
+        if constexpr (sizeof...(OutNames) > 0) {
+            constexpr std::array<std::string_view, sizeof...(OutNames)> names{OutNames.view()...};
+            return std::string("output['") + std::string(names[I]) + "']";
+        } else {
+            return "output[" + std::to_string(I) + "]";
+        }
+    }
+
+    // ── State ─────────────────────────────────────────────────────────────────
+
+    std::shared_ptr<IScheduler>              scheduler_;
+    std::string                              name_;
+    std::size_t                              fifo_capacity_;
+    input_channels_t                         input_channels_;
+    output_channels_t                        output_channels_{};
+    std::atomic<bool>                        stop_flag_{true};
+    std::atomic<bool>                        queued_{false};
+    NodeStats                                stats_;
+    NodeErrorHandler                         error_handler_;
+    std::chrono::milliseconds                max_exec_time_{0};
+    std::array<NodeEventCallback, 2>         event_callbacks_{};  // [0]=user [1]=network
+    std::array<NodeEventCallback, 2>         closed_callbacks_{};
+};
+
+// ── PoolObjectNode ────────────────────────────────────────────────────────────
+//
+// Same as PoolNode but wraps a stateful callable object (functor / class with
+// operator()). The object must outlive the PoolObjectNode.
+
+template<typename Obj,
+         typename InputTag  = in<>,
+         typename OutputTag = out<>,
+         fixed_string Label    = "",
+         std::size_t UniqueTag = 0>
+class PoolObjectNode;
+
+template<typename Obj, fixed_string... InNames, fixed_string... OutNames,
+         fixed_string Label, std::size_t UniqueTag>
+class PoolObjectNode<Obj, in<InNames...>, out<OutNames...>, Label, UniqueTag> : public INode {
+public:
+    using F            = decltype(&Obj::operator());
+    using args_tuple   = args_t<F>;
+    using return_raw   = return_t<F>;
+    using return_tuple = normalised_return_t<return_raw>;
+
+    static constexpr std::string_view label()    { return Label.view(); }
+    static constexpr std::size_t      unique_tag = UniqueTag;
+
+    static constexpr std::size_t input_count  = arity_v<F>;
+    static constexpr std::size_t output_count = std::tuple_size_v<return_tuple>;
+
+    static_assert(
+        sizeof...(InNames) == 0 || sizeof...(InNames) == input_count,
+        "make_pool_node: number of input names must match operator() arity, or provide none"
+    );
+    static_assert(
+        sizeof...(OutNames) == 0 || sizeof...(OutNames) == output_count,
+        "make_pool_node: number of output names must match return tuple size, or provide none"
+    );
+
+    explicit PoolObjectNode(Obj& obj, std::shared_ptr<IScheduler> sched,
+                            std::size_t fifo_capacity = 5)
+        : obj_(obj), scheduler_(std::move(sched)), fifo_capacity_(fifo_capacity)
+    {
+        init_input_channels(std::make_index_sequence<input_count>{});
+    }
+
+    ~PoolObjectNode() override { stop(); }
+
+    void start() override {
+        enable_inputs(std::make_index_sequence<input_count>{});
+        stop_flag_.store(false, std::memory_order_relaxed);
+        queued_.store(false, std::memory_order_relaxed);
+        register_callbacks(std::make_index_sequence<input_count>{});
+        if constexpr (input_count == 0)
+            try_submit(0.5f);
+    }
+
+    void stop() override {
+        stop_flag_.store(true, std::memory_order_seq_cst);
+        disable_inputs(std::make_index_sequence<input_count>{});
+    }
+
+    bool running() const override { return !stop_flag_.load(std::memory_order_relaxed); }
+    void set_name(std::string name) override { name_ = std::move(name); }
+    void set_error_handler(NodeErrorHandler h) { error_handler_ = std::move(h); }
+    void set_max_exec_time(std::chrono::milliseconds t) { max_exec_time_ = t; }
+
+    void set_overflow_callback(NodeEventCallback cb) { event_callbacks_[0] = std::move(cb); }
+    void set_network_overflow_callback(NodeEventCallback cb) override { event_callbacks_[1] = std::move(cb); }
+    void set_closed_callback(NodeEventCallback cb)   { closed_callbacks_[0] = std::move(cb); }
+    void set_network_closed_callback(NodeEventCallback cb)   override { closed_callbacks_[1] = std::move(cb); }
+
+    const NodeStats& stats() const override { return stats_; }
+
+    NodeSnapshot node_snapshot(const std::string& name, double elapsed_s) const override {
+        uint64_t frames   = stats_.frames_processed.load(std::memory_order_relaxed);
+        double exec_ms    = stats_.ema_exec_us.load(std::memory_order_relaxed) / 1000.0;
+        double blocked_ms = stats_.total_blocked_us.load(std::memory_order_relaxed) / 1000.0;
+        double qwait_ms   = stats_.queue_wait_us.load(std::memory_order_relaxed) / 1000.0;
+        double total_ms   = exec_ms + blocked_ms;
+        return {
+            name, frames, exec_ms,
+            stats_.max_exec_us.load(std::memory_order_relaxed) / 1000.0,
+            blocked_ms,
+            elapsed_s > 0 ? frames / elapsed_s : 0.0,
+            stats_.total_cpu_us.load(std::memory_order_relaxed) / 1000.0,
+            total_ms > 0 ? 100.0 * exec_ms / total_ms : 0.0,
+            qwait_ms,
+        };
+    }
+
+    template<std::size_t I> InputPort<PoolObjectNode, I>  input()  { return {*this}; }
+    template<std::size_t I> OutputPort<PoolObjectNode, I> output() { return {*this}; }
+
+    template<fixed_string Name>
+    auto input() {
+        constexpr std::size_t idx = index_of<Name, InNames...>();
+        static_assert(idx != npos, "unknown input port name");
+        return input<idx>();
+    }
+    template<fixed_string Name>
+    auto output() {
+        constexpr std::size_t idx = index_of<Name, OutNames...>();
+        static_assert(idx != npos, "unknown output port name");
+        return output<idx>();
+    }
+
+    template<std::size_t I>
+    Channel<std::tuple_element_t<I, args_tuple>>& input_channel() {
+        return *std::get<I>(input_channels_);
+    }
+    template<std::size_t I>
+    void set_input_channel(std::shared_ptr<Channel<std::tuple_element_t<I, args_tuple>>> ch) {
+        std::get<I>(input_channels_) = std::move(ch);
+    }
+    template<std::size_t I>
+    void set_output_channel(Channel<std::tuple_element_t<I, return_tuple>>* ch) {
+        std::get<I>(output_channels_) = ch;
+    }
+
+private:
+    template<std::size_t... Is>
+    void init_input_channels(std::index_sequence<Is...>) {
+        ((std::get<Is>(input_channels_) =
+              std::make_shared<Channel<std::tuple_element_t<Is, args_tuple>>>(fifo_capacity_)),
+         ...);
+    }
+    template<std::size_t... Is> void enable_inputs(std::index_sequence<Is...>)  { (std::get<Is>(input_channels_)->enable(), ...); }
+    template<std::size_t... Is> void disable_inputs(std::index_sequence<Is...>) { (std::get<Is>(input_channels_)->disable(), ...); }
+    template<std::size_t... Is>
+    void disable_outputs(std::index_sequence<Is...>) {
+        auto disable_one = [](auto* ch) { if (ch) ch->disable(); };
+        (disable_one(std::get<Is>(output_channels_)), ...);
+    }
+    template<std::size_t... Is>
+    void register_callbacks(std::index_sequence<Is...>) {
+        (std::get<Is>(input_channels_)->set_push_callback([this] { on_input_ready(); }), ...);
+    }
+
+    static void fire_callbacks(const std::array<NodeEventCallback, 2>& cbs) {
+        const auto ts = std::chrono::steady_clock::now();
+        for (auto& cb : cbs) if (cb) cb(ts);
+    }
+
+    void self_stop() {
+        disable_inputs(std::make_index_sequence<input_count>{});
+        disable_outputs(std::make_index_sequence<output_count>{});
+        stats_.exec_start_us.store(0, std::memory_order_relaxed);
+        queued_.store(false, std::memory_order_release);
+        stop_flag_.store(true, std::memory_order_relaxed);
+    }
+
+    template<typename Tup, std::size_t... Is>
+    static auto make_input_channel_tuple(std::index_sequence<Is...>)
+        -> std::tuple<std::shared_ptr<Channel<std::tuple_element_t<Is, Tup>>>...>;
+    using input_channels_t = decltype(make_input_channel_tuple<args_tuple>(
+        std::make_index_sequence<input_count>{}));
+
+    template<typename Tup, std::size_t... Is>
+    static auto make_output_channel_tuple(std::index_sequence<Is...>)
+        -> std::tuple<Channel<std::tuple_element_t<Is, Tup>>*...>;
+    using output_channels_t = decltype(make_output_channel_tuple<return_tuple>(
+        std::make_index_sequence<output_count>{}));
+
+    void on_input_ready() {
+        if (stop_flag_.load(std::memory_order_relaxed)) return;
+        std::size_t ready = count_ready(std::make_index_sequence<input_count>{});
+        if (ready == input_count) try_submit(compute_priority());
+    }
+
+    template<std::size_t... Is>
+    std::size_t count_ready(std::index_sequence<Is...>) {
+        return ((std::get<Is>(input_channels_)->approx_size() > 0 ? 1u : 0u) + ...);
+    }
+
+    float compute_priority() {
+        if constexpr (input_count == 0) return 0.5f;
+        float sum = 0.0f;
+        sum_fill(sum, std::make_index_sequence<input_count>{});
+        return sum / static_cast<float>(input_count);
+    }
+    template<std::size_t... Is>
+    void sum_fill(float& sum, std::index_sequence<Is...>) {
+        ((sum += std::get<Is>(input_channels_)->capacity() > 0
+            ? float(std::get<Is>(input_channels_)->approx_size())
+              / float(std::get<Is>(input_channels_)->capacity())
+            : 0.5f), ...);
+    }
+
+    void try_submit(float priority) {
+        bool expected = false;
+        if (queued_.compare_exchange_strong(expected, true, std::memory_order_acq_rel))
+            scheduler_->submit([this] { fire_once(); }, priority);
+    }
+
+    void fire_once() {
+        if (stop_flag_.load(std::memory_order_relaxed)) {
+            queued_.store(false, std::memory_order_release);
+            return;
+        }
+        auto t0 = clock_t::now();
+        int64_t now_us = std::chrono::duration_cast<std::chrono::microseconds>(
+            t0.time_since_epoch()).count();
+        stats_.exec_start_us.store(now_us, std::memory_order_relaxed);
+
+        try {
+            auto args = pop_inputs(std::make_index_sequence<input_count>{});
+            auto t1   = clock_t::now();
+            stats_.record_queue_wait(duration_t(t1 - t0));
+            auto cpu0 = NodeStats::cpu_now();
+
+            if constexpr (std::is_void_v<return_raw>) {
+                std::apply([this](auto&&... a) { obj_(std::forward<decltype(a)>(a)...); }, args);
+            } else {
+                auto result = std::apply([this](auto&&... a) { return obj_(std::forward<decltype(a)>(a)...); }, args);
+                push_outputs(normalise(std::move(result)), std::make_index_sequence<output_count>{});
+            }
+
+            auto cpu1 = NodeStats::cpu_now();
+            auto t2   = clock_t::now();
+            stats_.record_exec(duration_t(t2 - t1), duration_t::zero(), cpu0, cpu1);
+        } catch (const ChannelClosedError&) {
+            fire_callbacks(closed_callbacks_);
+            self_stop();
+            return;
+        } catch (const ChannelOverflowError&) {
+            fire_callbacks(event_callbacks_);
+        } catch (...) {
+            if (error_handler_ && error_handler_(name_, std::current_exception())) {
+            } else {
+                fire_callbacks(closed_callbacks_);
+                self_stop();
+                return;
+            }
+        }
+
+        stats_.exec_start_us.store(0, std::memory_order_relaxed);
+        queued_.store(false, std::memory_order_release);
+        if (stop_flag_.load(std::memory_order_relaxed)) return;
+        if constexpr (input_count == 0) try_submit(0.5f);
+        else                            on_input_ready();
+    }
+
+    template<std::size_t... Is>
+    args_tuple pop_inputs(std::index_sequence<Is...>) { return {pop_one<Is>()...}; }
+
+    template<std::size_t I>
+    std::tuple_element_t<I, args_tuple> pop_one() {
+        auto& ch = *std::get<I>(input_channels_);
+        std::tuple_element_t<I, args_tuple> val;
+        if (!ch.try_pop_now(val)) throw ChannelClosedError{};
+        return val;
+    }
+
+    template<typename R = return_raw>
+    static return_tuple normalise(R&& r) {
+        if constexpr (is_tuple_v<R>) return std::move(r);
+        else                         return std::make_tuple(std::move(r));
+    }
+
+    template<std::size_t... Is>
+    void push_outputs(return_tuple&& result, std::index_sequence<Is...>) {
+        (push_one_out<Is>(std::get<Is>(std::move(result))), ...);
+    }
+    template<std::size_t I>
+    void push_one_out(std::tuple_element_t<I, return_tuple>&& val) {
+        auto* ch = std::get<I>(output_channels_);
+        if (!ch) return;
+        try {
+            ch->push(std::move(val));
+        } catch (const ChannelOverflowError&) {
+            throw ChannelOverflowError(ch->capacity(),
+                "pool node '" + name_ + "'");
+        }
+    }
+
+    Obj&                                     obj_;
+    std::shared_ptr<IScheduler>              scheduler_;
+    std::string                              name_;
+    std::size_t                              fifo_capacity_;
+    input_channels_t                         input_channels_;
+    output_channels_t                        output_channels_{};
+    std::atomic<bool>                        stop_flag_{true};
+    std::atomic<bool>                        queued_{false};
+    NodeStats                                stats_;
+    NodeErrorHandler                         error_handler_;
+    std::chrono::milliseconds                max_exec_time_{0};
+    std::array<NodeEventCallback, 2>         event_callbacks_{};  // [0]=user [1]=network
+    std::array<NodeEventCallback, 2>         closed_callbacks_{};
+};
+
+// ── make_pool_node factory (NTTP) ─────────────────────────────────────────────
+
+template<auto Func, fixed_string Label = "", std::size_t UniqueTag = 0>
+auto make_pool_node(std::shared_ptr<IScheduler> sched, std::size_t fifo_capacity = 5) {
+    return PoolNode<Func, in<>, out<>, Label, UniqueTag>(std::move(sched), fifo_capacity);
+}
+
+template<auto Func, fixed_string Label = "", std::size_t UniqueTag = 0,
+         fixed_string... InNames>
+auto make_pool_node(std::shared_ptr<IScheduler> sched, in<InNames...>,
+                    std::size_t fifo_capacity = 5) {
+    return PoolNode<Func, in<InNames...>, out<>, Label, UniqueTag>(std::move(sched), fifo_capacity);
+}
+
+template<auto Func, fixed_string Label = "", std::size_t UniqueTag = 0,
+         fixed_string... OutNames>
+auto make_pool_node(std::shared_ptr<IScheduler> sched, out<OutNames...>,
+                    std::size_t fifo_capacity = 5) {
+    return PoolNode<Func, in<>, out<OutNames...>, Label, UniqueTag>(std::move(sched), fifo_capacity);
+}
+
+template<auto Func, fixed_string Label = "", std::size_t UniqueTag = 0,
+         fixed_string... InNames, fixed_string... OutNames>
+auto make_pool_node(std::shared_ptr<IScheduler> sched, in<InNames...>, out<OutNames...>,
+                    std::size_t fifo_capacity = 5) {
+    return PoolNode<Func, in<InNames...>, out<OutNames...>, Label, UniqueTag>(
+        std::move(sched), fifo_capacity);
+}
+
+// ── make_pool_node factory (callable object) ──────────────────────────────────
+
+template<typename Obj>
+auto make_pool_node(Obj& obj, std::shared_ptr<IScheduler> sched,
+                    std::size_t fifo_capacity = 5) {
+    return PoolObjectNode<Obj, in<>, out<>>(obj, std::move(sched), fifo_capacity);
+}
+
+template<typename Obj, fixed_string... InNames>
+auto make_pool_node(Obj& obj, std::shared_ptr<IScheduler> sched, in<InNames...>,
+                    std::size_t fifo_capacity = 5) {
+    return PoolObjectNode<Obj, in<InNames...>, out<>>(obj, std::move(sched), fifo_capacity);
+}
+
+template<typename Obj, fixed_string... OutNames>
+auto make_pool_node(Obj& obj, std::shared_ptr<IScheduler> sched, out<OutNames...>,
+                    std::size_t fifo_capacity = 5) {
+    return PoolObjectNode<Obj, in<>, out<OutNames...>>(obj, std::move(sched), fifo_capacity);
+}
+
+template<typename Obj, fixed_string... InNames, fixed_string... OutNames>
+auto make_pool_node(Obj& obj, std::shared_ptr<IScheduler> sched,
+                    in<InNames...>, out<OutNames...>,
+                    std::size_t fifo_capacity = 5) {
+    return PoolObjectNode<Obj, in<InNames...>, out<OutNames...>>(
+        obj, std::move(sched), fifo_capacity);
+}
+
+} // namespace kpn

include/kpn/python/auto_bind.hpp

@@ -0,0 +1,311 @@
+#pragma once
+// Auto-binding helpers for KPN++ Python bindings.
+//
+// Usage in your binding .cpp:
+//
+//   #define KPN_BUILD_PYTHON
+//   #include <kpn/python/auto_bind.hpp>
+//
+//   int produce()        { return 42; }
+//   int double_it(int x) { return x * 2; }
+//   void print_it(int x) { std::cout << x << '\n'; }
+//
+//   using MyNodes = kpn::python::NodeRegistry<
+//       kpn::python::Entry<produce,   "produce">,
+//       kpn::python::Entry<double_it, "double_it">,
+//       kpn::python::Entry<print_it,  "print_it">
+//   >;
+//
+//   NB_MODULE(my_kpn, m) {
+//       kpn::python::bind_network<MyNodes>(m);   // KPN_BIND_PYTHON behaviour
+//       kpn::python::bind_debug<MyNodes>(m);     // KPN_PYTHON_DEBUG behaviour
+//   }
+//
+// bind_network registers:
+//   - Network class (PyNetwork<auto-deduced-variant>) with auto-registered converters
+//   - make_<name>(capacity=5) factory for each entry
+//   - <Name>Node class for each entry
+//
+// bind_debug additionally registers each raw C++ function as a free Python
+// callable (e.g. double_it(5) → 10) so node logic can be tested without a network.
+//
+// To support a custom type T, specialise kpn::PythonConverter<T> before calling
+// bind_network:
+//
+//   namespace kpn {
+//   template<> struct PythonConverter<MyVec3> {
+//       static constexpr const char* type_name = "vec3";  // optional friendly name
+//       static nb::object to_python(const MyVec3& v) { ... }
+//       static MyVec3 from_python(nb::object o)      { ... }
+//   };
+//   } // namespace kpn
+
+#include "../variant_node.hpp"
+#include "../traits.hpp"
+#include "bindings.hpp"
+
+#ifdef KPN_BUILD_PYTHON
+#include <nanobind/nanobind.h>
+#include <nanobind/stl/shared_ptr.h>
+#include <nanobind/stl/string.h>
+#include <nanobind/stl/vector.h>
+
+#include <cctype>
+#include <string>
+#include <tuple>
+#include <type_traits>
+
+// ── PythonConverter specialisations for built-in nanobind-castable types ─────
+// These live in kpn:: to match the primary template in variant_node.hpp.
+
+namespace kpn {
+
+template<> struct PythonConverter<int> {
+    static constexpr const char* type_name = "int";
+    static nanobind::object to_python(const int& v) { return nanobind::cast(v); }
+    static int from_python(nanobind::object o) { return nanobind::cast<int>(std::move(o)); }
+};
+
+template<> struct PythonConverter<float> {
+    static constexpr const char* type_name = "float";
+    static nanobind::object to_python(const float& v) { return nanobind::cast(v); }
+    static float from_python(nanobind::object o) { return nanobind::cast<float>(std::move(o)); }
+};
+
+template<> struct PythonConverter<double> {
+    static constexpr const char* type_name = "double";
+    static nanobind::object to_python(const double& v) { return nanobind::cast(v); }
+    static double from_python(nanobind::object o) { return nanobind::cast<double>(std::move(o)); }
+};
+
+template<> struct PythonConverter<bool> {
+    static constexpr const char* type_name = "bool";
+    static nanobind::object to_python(const bool& v) { return nanobind::cast(v); }
+    static bool from_python(nanobind::object o) { return nanobind::cast<bool>(std::move(o)); }
+};
+
+template<> struct PythonConverter<std::string> {
+    static constexpr const char* type_name = "str";
+    static nanobind::object to_python(const std::string& v) { return nanobind::cast(v); }
+    static std::string from_python(nanobind::object o) {
+        return nanobind::cast<std::string>(std::move(o));
+    }
+};
+
+} // namespace kpn
+
+namespace kpn::python {
+namespace nb = nanobind;
+
+// ── Entry<Func, Name> ─────────────────────────────────────────────────────────
+// Compile-time descriptor for one bindable node function.
+
+template<auto Func, fixed_string Name>
+struct Entry {
+    static constexpr auto func = Func;
+    static constexpr auto name = Name;
+};
+
+// ── NodeRegistry<Es...> ───────────────────────────────────────────────────────
+
+template<typename... Es>
+struct NodeRegistry {
+    using entries_tuple = std::tuple<Es...>;
+    static constexpr std::size_t size = sizeof...(Es);
+};
+
+// ── Internal TMP ──────────────────────────────────────────────────────────────
+
+namespace detail {
+
+// All non-void port types for a single function (args + normalised returns).
+template<auto Func>
+struct entry_port_types {
+    using args = args_t<decltype(Func)>;
+    using ret  = normalised_return_t<return_t<decltype(Func)>>;
+    using type = decltype(std::tuple_cat(std::declval<args>(), std::declval<ret>()));
+};
+
+// Flat tuple of all types across all entries (may contain duplicates).
+template<typename... Es>
+struct all_types_flat {
+    using type = decltype(std::tuple_cat(
+        std::declval<typename entry_port_types<Es::func>::type>()...));
+};
+
+template<typename Registry>
+struct registry_flat_types;
+
+template<typename... Es>
+struct registry_flat_types<NodeRegistry<Es...>> {
+    using type = typename all_types_flat<Es...>::type;
+};
+
+// Unpack a tuple into unique_types_t (which takes a pack, not a tuple).
+// unique_types_t<T> takes Ts... not std::tuple<Ts...>, so we need this bridge.
+template<typename Tuple>
+struct unpack_unique;
+
+template<typename... Ts>
+struct unpack_unique<std::tuple<Ts...>> {
+    using type = kpn::detail::unique_types_t<Ts...>;
+};
+
+// SFINAE: does PythonConverter<T> have a 'type_name' member?
+template<typename Conv, typename = void>
+struct has_type_name : std::false_type {};
+
+template<typename Conv>
+struct has_type_name<Conv, std::void_t<decltype(Conv::type_name)>> : std::true_type {};
+
+inline std::string make_class_name(std::string_view snake) {
+    std::string result(snake);
+    if (!result.empty()) result[0] = static_cast<char>(std::toupper(result[0]));
+    result += "Node";
+    return result;
+}
+
+} // namespace detail
+
+// ── registry_variant_t<Registry> ─────────────────────────────────────────────
+// Deduces std::variant<UniqueTypes...> from all port types across the registry.
+
+template<typename Registry>
+using registry_variant_t = typename kpn::detail::tuple_to_variant<
+    typename detail::unpack_unique<
+        typename detail::registry_flat_types<Registry>::type>::type
+>::type;
+
+// ── Converter registration ────────────────────────────────────────────────────
+
+template<typename T, typename Variant>
+void register_one_type(PyNetwork<Variant>& net) {
+    const char* friendly = nullptr;
+    if constexpr (detail::has_type_name<PythonConverter<T>>::value)
+        friendly = PythonConverter<T>::type_name;
+
+    net.template register_full_type<T>(
+        [](const T& v) -> nb::object { return PythonConverter<T>::to_python(v); },
+        [](nb::object o) -> T        { return PythonConverter<T>::from_python(std::move(o)); },
+        friendly);
+}
+
+template<typename Variant, typename... Ts>
+void register_types_impl(PyNetwork<Variant>& net, std::tuple<Ts...>*) {
+    (register_one_type<Ts>(net), ...);
+}
+
+template<typename Registry, typename Variant>
+void register_all_converters(PyNetwork<Variant>& net) {
+    using Flat   = typename detail::registry_flat_types<Registry>::type;
+    using Unique = typename detail::unpack_unique<Flat>::type;
+    register_types_impl(net, static_cast<Unique*>(nullptr));
+}
+
+// ── Per-entry class + factory registration ───────────────────────────────────
+
+namespace detail {
+
+template<typename E, typename Variant>
+void register_one_entry(nb::module_& m) {
+    using Wrapper = VariantNodeWrapper<E::func, Variant>;
+
+    auto class_name = make_class_name(E::name.view());
+    auto make_name  = "make_" + std::string(E::name.view());
+
+    nb::class_<Wrapper, IVariantNode<Variant>>(m, class_name.c_str())
+        .def("__init__", [](Wrapper* self, std::size_t cap) {
+            new (self) Wrapper(cap);
+        }, nb::arg("capacity") = 5);
+
+    m.def(make_name.c_str(),
+          [](std::size_t cap) -> std::shared_ptr<IVariantNode<Variant>> {
+              return std::make_shared<Wrapper>(cap);
+          },
+          nb::arg("capacity") = 5);
+}
+
+template<typename Variant, typename... Es>
+void register_entries_impl(nb::module_& m, std::tuple<Es...>*) {
+    (register_one_entry<Es, Variant>(m), ...);
+}
+
+} // namespace detail
+
+// ── bind_network<Registry> ────────────────────────────────────────────────────
+// Registers:
+//   - INode          — base class (opaque Python handle)
+//   - Network        — PyNetwork with auto-registered converters
+//   - <Name>Node     — VariantNodeWrapper for each entry
+//   - make_<name>()  — factory returning shared_ptr<INode>
+
+template<typename Registry>
+void bind_network(nb::module_& m) {
+    using Variant = registry_variant_t<Registry>;
+    using Net     = PyNetwork<Variant>;
+    using Entries = typename Registry::entries_tuple;
+
+    nb::class_<IVariantNode<Variant>>(m, "INode");
+
+    nb::class_<Net>(m, "Network")
+        .def("__init__", [](Net* self) {
+            new (self) Net();
+            register_all_converters<Registry>(*self);
+        })
+        // add(name, c++_node)
+        .def("add", [](Net& self, std::string name,
+                        std::shared_ptr<IVariantNode<Variant>> node) {
+            self.add(std::move(name), std::move(node));
+        }, nb::arg("name"), nb::arg("node"))
+        // add_node(name, callable, inputs=[...], outputs=[...], capacity=5)
+        .def("add_node", &Net::add_node_python,
+             nb::arg("name"),
+             nb::arg("callable"),
+             nb::arg("inputs")   = std::vector<std::string>{},
+             nb::arg("outputs")  = std::vector<std::string>{},
+             nb::arg("capacity") = std::size_t(5))
+        .def("connect", &Net::connect,
+             nb::arg("src"), nb::arg("out_idx"),
+             nb::arg("dst"), nb::arg("in_idx"))
+        .def("build",  &Net::build)
+        .def("start",  &Net::start)
+        .def("stop",   &Net::stop)
+        .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"))
+    ;
+
+    detail::register_entries_impl<Variant>(m, static_cast<Entries*>(nullptr));
+}
+
+// ── bind_debug<Registry> ─────────────────────────────────────────────────────
+// Exposes each node's raw C++ function as a free Python callable so node logic
+// can be unit-tested without constructing a network.
+//
+// Example:  assert kpn.double_it(5) == 10
+
+namespace detail {
+
+template<typename E>
+void bind_one_debug(nb::module_& m) {
+    auto name_str = std::string(E::name.view());
+    m.def(name_str.c_str(), E::func);
+}
+
+template<typename... Es>
+void bind_debug_impl(nb::module_& m, std::tuple<Es...>*) {
+    (bind_one_debug<Es>(m), ...);
+}
+
+} // namespace detail
+
+template<typename Registry>
+void bind_debug(nb::module_& m) {
+    using Entries = typename Registry::entries_tuple;
+    detail::bind_debug_impl(m, static_cast<Entries*>(nullptr));
+}
+
+} // namespace kpn::python
+
+#endif // KPN_BUILD_PYTHON

include/kpn/python/bindings.hpp

@@ -26,6 +26,9 @@ namespace nb = nanobind;
 // them via IVariantChannel adapters. The variant only lives at the boundary;
 // each node's internal Channel<T> stores raw T values.
 
+template<typename Variant>
+class PyNode; // forward declaration
+
 template<typename Variant>
 class PyNetwork {
 public:
@@ -43,8 +46,6 @@ public:
     }
 
     // connect(src_name, out_idx, dst_name, in_idx)
-    // Wires src's output port out_idx to dst's input port in_idx.
-    // Type check: both sides must carry the same T.
     void connect(const std::string& src_name, std::size_t out_idx,
                  const std::string& dst_name, std::size_t in_idx)
     {
@@ -65,7 +66,6 @@ public:
                 dst_name + ".input[" + std::to_string(in_idx) +
                 "] (" + dst.input_type(in_idx).name() + ")");
 
-        // The destination node owns the input channel — get it, then tell src to use it.
         auto ch = dst.input_channel(in_idx);
         src.set_output_channel(out_idx, std::move(ch));
         adj_[src_name].push_back(dst_name);
@@ -92,16 +92,12 @@ public:
 
     // ── Python tap/inject ─────────────────────────────────────────────────────
 
-    // Read one value from node's output port. Releases GIL while blocking.
     nb::object read(const std::string& node_name, std::size_t out_idx) {
-        // We need a channel that sits on the output of this node.
-        // read() installs a tap channel if not already present.
         auto key = tap_key(node_name, out_idx);
         if (!taps_.count(key)) {
             auto& src = node_at(node_name);
             if (out_idx >= src.output_count())
                 throw std::out_of_range(node_name + ": output index out of range");
-            // Create a tap channel matching the output type and wire it
             auto tap = make_tap_channel(src.output_type(out_idx));
             src.set_output_channel(out_idx, tap);
             taps_[key] = std::move(tap);
@@ -114,7 +110,6 @@ public:
         return variant_to_python(std::move(v));
     }
 
-    // Write a Python value into node's input port. Releases GIL while blocking.
     void write(const std::string& node_name, std::size_t in_idx, nb::object value) {
         auto& dst = node_at(node_name);
         if (in_idx >= dst.input_count())
@@ -128,8 +123,31 @@ public:
         }
     }
 
-    // ── Converter registration ────────────────────────────────────────────────
-    // Called once per type at module init time to register to/from Python converters.
+    // ── Python-callable node creation ─────────────────────────────────────────
+    // Creates a PyNode wrapping a Python callable and adds it to the graph.
+    // Type names must have been registered via register_full_type<T>().
+
+    void add_node_python(std::string name, nb::object callable,
+                         std::vector<std::string> in_names,
+                         std::vector<std::string> out_names,
+                         std::size_t capacity = 5)
+    {
+        std::vector<std::type_index> in_types, out_types;
+        for (auto& s : in_names)  in_types.push_back(resolve_type_name(s));
+        for (auto& s : out_names) out_types.push_back(resolve_type_name(s));
+
+        add(std::move(name),
+            std::make_shared<PyNode<Variant>>(
+                std::move(callable),
+                std::move(in_types),
+                std::move(out_types),
+                to_python_,
+                from_python_,
+                ch_factories_,
+                capacity));
+    }
+
+    // ── Type converter registration ───────────────────────────────────────────
 
     template<typename T>
     void register_type(
@@ -143,6 +161,52 @@ public:
         };
     }
 
+    // register_channel_factory<T>: registers factory for creating input channels.
+    template<typename T>
+    void register_channel_factory() {
+        ch_factories_[std::type_index(typeid(T))] =
+            [](std::size_t cap) -> std::shared_ptr<VChannel> {
+                return std::make_shared<VariantChannel<T, Variant>>(
+                    std::make_shared<Channel<T>>(cap));
+            };
+    }
+
+    // Backward-compatible alias.
+    template<typename T>
+    void register_tap_factory(std::size_t = 5) {
+        register_channel_factory<T>();
+    }
+
+    // register_full_type<T>: registers converters + channel factory + type name.
+    // This is what auto_bind.hpp calls; manual bindings can call register_type +
+    // register_tap_factory separately for backward compatibility.
+    template<typename T>
+    void register_full_type(
+        std::function<nb::object(const T&)> to_py,
+        std::function<T(nb::object)>        from_py,
+        const char* friendly_name = nullptr)
+    {
+        register_type<T>(std::move(to_py), std::move(from_py));
+        register_channel_factory<T>();
+        auto idx = std::type_index(typeid(T));
+        type_names_.insert_or_assign(typeid(T).name(), idx);
+        if (friendly_name) type_names_.insert_or_assign(friendly_name, idx);
+    }
+
+    // ── Type name lookup ──────────────────────────────────────────────────────
+
+    void register_type_name(const std::string& name, std::type_index idx) {
+        type_names_.insert_or_assign(name, idx);
+    }
+
+    std::type_index resolve_type_name(const std::string& name) const {
+        auto it = type_names_.find(name);
+        if (it == type_names_.end())
+            throw std::runtime_error(
+                "type '" + name + "' not registered — call register_full_type<T>() first");
+        return it->second;
+    }
+
 private:
     VNode& node_at(const std::string& name) {
         auto it = nodes_.find(name);
@@ -166,15 +230,14 @@ private:
         return node + ":" + std::to_string(idx);
     }
 
-    std::shared_ptr<VChannel> make_tap_channel(std::type_index type) {
-        // Create the right VariantChannel<T> based on the registered type index.
-        // We need a factory registered per type — stored in tap_factories_.
-        auto it = tap_factories_.find(type);
-        if (it == tap_factories_.end())
+    std::shared_ptr<VChannel> make_tap_channel(std::type_index type,
+                                                std::size_t cap = 5) {
+        auto it = ch_factories_.find(type);
+        if (it == ch_factories_.end())
             throw std::runtime_error(
-                "no tap factory for type: " + std::string(type.name()) +
-                " — was register_type() called for this type?");
-        return it->second();
+                "no channel factory for type: " + std::string(type.name()) +
+                " — call register_full_type<T>() or register_tap_factory<T>()");
+        return it->second(cap);
     }
 
     nb::object variant_to_python(Variant v) {
@@ -194,18 +257,6 @@ private:
         return it->second(std::move(obj));
     }
 
-public:
-    // Called by register_type to also register a tap channel factory.
-    template<typename T>
-    void register_tap_factory(std::size_t capacity = 5) {
-        auto idx = std::type_index(typeid(T));
-        tap_factories_[idx] = [capacity]() -> std::shared_ptr<VChannel> {
-            auto ch = std::make_shared<Channel<T>>(capacity);
-            return std::make_shared<VariantChannel<T, Variant>>(std::move(ch));
-        };
-    }
-
-private:
     std::map<std::string, std::shared_ptr<VNode>>    nodes_;
     std::map<std::string, std::vector<std::string>>  adj_;
     std::vector<std::string>                         topo_;
@@ -213,19 +264,27 @@ private:
 
     std::map<std::type_index, std::function<nb::object(const Variant&)>> to_python_;
     std::map<std::type_index, std::function<Variant(nb::object)>>        from_python_;
-    std::map<std::type_index, std::function<std::shared_ptr<VChannel>()>> tap_factories_;
+
+    // Channel factory: type → function(capacity) → VChannel.
+    // Used both for tap channels (read()) and PyNode input channel creation.
+    std::map<std::type_index,
+             std::function<std::shared_ptr<VChannel>(std::size_t)>> ch_factories_;
+
+    // Friendly name → type_index (e.g. "int" → typeid(int)).
+    std::map<std::string, std::type_index> type_names_;
 };
 
 // ── PyNode<Variant> ───────────────────────────────────────────────────────────
 // A pure-Python processing node. Holds a nanobind callable.
-// run_loop: pop inputs (release GIL), call Python (acquire GIL), push outputs (release GIL).
+// run_loop: pop inputs (release GIL), call Python (acquire GIL), push outputs.
 
 template<typename Variant>
 class PyNode : public IVariantNode<Variant> {
 public:
     using VChannel = IVariantChannel<Variant>;
 
-    using ChannelFactory = std::function<std::shared_ptr<VChannel>(std::size_t capacity)>;
+    using ChannelFactory =
+        std::function<std::shared_ptr<VChannel>(std::size_t capacity)>;
 
     PyNode(nb::object callable,
            std::vector<std::type_index> in_types,
@@ -264,7 +323,6 @@ public:
         for (auto& ch : in_channels_) ch->disable();
         if (thread_.joinable()) {
             thread_.request_stop();
-            // Release GIL while joining — run_loop may be waiting to acquire it.
             nb::gil_scoped_release release;
             thread_.join();
         }
@@ -311,20 +369,17 @@ public:
 
 private:
     void run_loop() {
-        // This thread does not hold the GIL. It acquires it only for Python calls.
         while (!stop_flag_.load(std::memory_order_relaxed)) {
             try {
                 auto t0 = clock_t::now();
 
-                // Pop all inputs — no GIL needed, these are pure C++ channel ops
                 std::vector<Variant> inputs(in_channels_.size());
                 for (std::size_t i = 0; i < in_channels_.size(); ++i)
                     inputs[i] = in_channels_[i]->pop();
 
-                auto t1 = clock_t::now();
+                auto t1   = clock_t::now();
                 auto cpu0 = NodeStats::cpu_now();
 
-                // Acquire GIL only for the Python call and type conversion
                 std::vector<Variant> outputs;
                 {
                     nb::gil_scoped_acquire acquire;
@@ -344,10 +399,9 @@ private:
                 }
 
                 auto cpu1 = NodeStats::cpu_now();
-                auto t2 = clock_t::now();
+                auto t2   = clock_t::now();
                 stats_.record_exec(duration_t(t2 - t1), duration_t(t1 - t0), cpu0, cpu1);
 
-                // Push outputs — no GIL needed
                 for (std::size_t i = 0; i < out_channels_.size(); ++i) {
                     if (out_channels_[i])
                         out_channels_[i]->push(std::move(outputs[i]));
@@ -385,9 +439,9 @@ private:
     NodeStats          stats_;
 };
 
-// ── register_py_network ───────────────────────────────────────────────────────
-// Registers PyNetwork<Variant> and PyNode<Variant> with the given nanobind module.
-// Call once per module, passing the Variant type derived from your registered node types.
+// ── register_py_network (legacy helper) ───────────────────────────────────────
+// Registers PyNetwork<Variant> with the given nanobind module.
+// Prefer bind_network<Registry> from auto_bind.hpp for new code.
 
 template<typename Variant>
 void register_py_network(nb::module_& m, const char* class_name = "Network") {
@@ -402,7 +456,7 @@ void register_py_network(nb::module_& m, const char* class_name = "Network") {
         .def("start",  &Net::start)
         .def("stop",   &Net::stop)
         .def("read",   &Net::read,
-             nb::arg("node"), nb::arg("out_idx") = 0)
+             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"));
 }

include/kpn/scheduler.hpp

@@ -0,0 +1,211 @@
+#pragma once
+#include "diagnostics.hpp"
+#include <atomic>
+#include <condition_variable>
+#include <functional>
+#include <memory>
+#include <mutex>
+#include <optional>
+#include <queue>
+#include <thread>
+#include <vector>
+
+namespace kpn {
+
+// ── IScheduler ────────────────────────────────────────────────────────────────
+
+struct IScheduler {
+    virtual ~IScheduler() = default;
+
+    // Submit a task with an optional priority in [0, 1]. Higher = run sooner.
+    virtual void submit(std::function<void()> task, float priority = 0.5f) = 0;
+
+    // Start worker threads. Must be called before submit().
+    virtual void start() = 0;
+
+    // Halt: signal workers to exit and join them. Pending tasks are discarded.
+    virtual void stop() = 0;
+
+    // Drain: block until all in-flight tasks complete. Workers keep running.
+    virtual void drain() = 0;
+};
+
+// ── ThreadPool ────────────────────────────────────────────────────────────────
+//
+// Work-stealing thread pool with per-thread priority queues.
+//
+// Each worker owns a priority_queue (max-heap by priority, FIFO within equal
+// priority via sequence number). submit() distributes via round-robin. When a
+// worker's queue is empty it tries to steal from the most-loaded peer using
+// try_lock to avoid blocking; if no work is found it sleeps on a shared CV.
+//
+// total_ counts tasks submitted-but-not-completed (queued + executing).
+// drain() waits until total_ == 0.
+
+class ThreadPool : public IScheduler, public IPoolProbe {
+public:
+    explicit ThreadPool(std::size_t thread_count) : thread_count_(thread_count) {}
+
+    ~ThreadPool() {
+        if (!stopped_.load(std::memory_order_relaxed))
+            stop();
+    }
+
+    void start() override {
+        stopped_.store(false, std::memory_order_relaxed);
+        queues_.clear();
+        for (std::size_t i = 0; i < thread_count_; ++i)
+            queues_.push_back(std::make_unique<WorkerQueue>());
+        workers_.reserve(thread_count_);
+        for (std::size_t i = 0; i < thread_count_; ++i)
+            workers_.emplace_back([this, i] { worker_loop(i); });
+    }
+
+    void stop() override {
+        stopped_.store(true, std::memory_order_seq_cst);
+        for (auto& q : queues_) {
+            std::lock_guard lock(q->mx);
+            std::size_t discarded = q->pq.size();
+            while (!q->pq.empty()) q->pq.pop();
+            total_.fetch_sub(discarded, std::memory_order_relaxed);
+        }
+        // Lock cv_mx_ before notifying so the stop signal can't be lost in the
+        // gap between a worker's predicate check and its wait() (see submit()).
+        { std::lock_guard<std::mutex> lk(cv_mx_); }
+        cv_.notify_all();
+        for (auto& t : workers_) if (t.joinable()) t.join();
+        workers_.clear();
+        queues_.clear();
+    }
+
+    void drain() override {
+        std::unique_lock lock(drain_mx_);
+        drain_cv_.wait(lock, [this] {
+            return total_.load(std::memory_order_acquire) == 0;
+        });
+    }
+
+    void submit(std::function<void()> task, float priority = 0.5f) override {
+        std::size_t target = next_.fetch_add(1, std::memory_order_relaxed) % thread_count_;
+        {
+            std::lock_guard lock(queues_[target]->mx);
+            queues_[target]->pq.push(
+                {std::move(task), priority, seq_.fetch_add(1, std::memory_order_relaxed)});
+        }
+        total_.fetch_add(1, std::memory_order_relaxed);
+        submitted_.fetch_add(1, std::memory_order_relaxed);
+        // Synchronize with worker_loop's predicate evaluation: taking cv_mx_
+        // here guarantees a worker is either before its predicate check (and
+        // will observe total_ > 0) or already blocked in wait() (and will be
+        // woken). Without this, notify_one() can slip into the gap between the
+        // worker's predicate check and its wait(), and be lost — a deadlock.
+        { std::lock_guard<std::mutex> lk(cv_mx_); }
+        cv_.notify_one();
+    }
+
+    std::size_t thread_count() const { return thread_count_; }
+
+    // ── IPoolProbe ────────────────────────────────────────────────────────────
+
+    PoolSnapshot snapshot(const std::string& name) const override {
+        std::size_t a = active_.load(std::memory_order_relaxed);
+        std::size_t t = total_.load(std::memory_order_relaxed);
+        return {
+            name, thread_count_,
+            t > a ? t - a : 0,   // queued (approximate)
+            a,                    // executing
+            submitted_.load(std::memory_order_relaxed),
+            completed_.load(std::memory_order_relaxed),
+        };
+    }
+
+private:
+    struct Task {
+        std::function<void()> fn;
+        float    priority;
+        uint64_t seq;
+        // max-heap: higher priority runs first; older task wins tie
+        bool operator<(const Task& o) const {
+            if (priority != o.priority) return priority < o.priority;
+            return seq > o.seq;
+        }
+    };
+
+    // Separate cache lines to prevent false sharing between adjacent queues.
+    struct alignas(64) WorkerQueue {
+        std::priority_queue<Task> pq;
+        std::mutex                mx;
+    };
+
+    std::optional<std::function<void()>> try_pop(WorkerQueue& q) {
+        std::lock_guard lock(q.mx);
+        if (q.pq.empty()) return std::nullopt;
+        auto fn = std::move(const_cast<Task&>(q.pq.top()).fn);
+        q.pq.pop();
+        return fn;
+    }
+
+    std::optional<std::function<void()>> try_steal(std::size_t thief) {
+        // Find the most-loaded peer without blocking — racy peek is fine.
+        std::size_t victim = thief, best = 0;
+        for (std::size_t i = 0; i < queues_.size(); ++i) {
+            if (i == thief) continue;
+            std::unique_lock lk(queues_[i]->mx, std::try_to_lock);
+            if (!lk) continue;
+            std::size_t n = queues_[i]->pq.size();
+            if (n > best) { best = n; victim = i; }
+        }
+        if (victim == thief) return std::nullopt;
+        return try_pop(*queues_[victim]);
+    }
+
+    void execute(std::function<void()>& fn) {
+        active_.fetch_add(1, std::memory_order_relaxed);
+        fn();
+        completed_.fetch_add(1, std::memory_order_relaxed);
+        active_.fetch_sub(1, std::memory_order_relaxed);
+        // Notify drain() if this was the last in-flight task.
+        // acq_rel ensures the decrement is visible before any drain() load.
+        // Lock drain_mx_ before notifying to avoid a lost wakeup against
+        // drain()'s predicate check (same hazard as submit()/cv_mx_).
+        if (total_.fetch_sub(1, std::memory_order_acq_rel) == 1) {
+            { std::lock_guard<std::mutex> lk(drain_mx_); }
+            drain_cv_.notify_all();
+        }
+    }
+
+    void worker_loop(std::size_t id) {
+        while (true) {
+            if (auto fn = try_pop(*queues_[id]))   { execute(*fn); continue; }
+            if (auto fn = try_steal(id))            { execute(*fn); continue; }
+
+            std::unique_lock lock(cv_mx_);
+            cv_.wait(lock, [this] {
+                return stopped_.load(std::memory_order_seq_cst)
+                    || total_.load(std::memory_order_relaxed) > 0;
+            });
+            if (stopped_.load(std::memory_order_seq_cst)
+                && total_.load(std::memory_order_relaxed) == 0)
+                return;
+        }
+    }
+
+    const std::size_t                          thread_count_;
+    std::vector<std::unique_ptr<WorkerQueue>>  queues_;
+    std::vector<std::thread>                   workers_;
+
+    std::mutex              cv_mx_;
+    std::condition_variable cv_;
+    std::mutex              drain_mx_;
+    std::condition_variable drain_cv_;
+
+    std::atomic<bool>     stopped_{true};
+    std::atomic<size_t>   total_{0};      // queued + executing
+    std::atomic<size_t>   active_{0};     // executing only (for snapshot)
+    std::atomic<size_t>   next_{0};       // round-robin submit cursor
+    std::atomic<uint64_t> seq_{0};        // tie-break for equal-priority tasks
+    std::atomic<uint64_t> submitted_{0};
+    std::atomic<uint64_t> completed_{0};
+};
+
+} // namespace kpn

include/kpn/shared_resource.hpp

@@ -0,0 +1,190 @@
+#pragma once
+#include "diagnostics.hpp"
+
+#include <algorithm>
+#include <atomic>
+#include <chrono>
+#include <condition_variable>
+#include <functional>
+#include <mutex>
+#include <utility>
+#include <vector>
+
+namespace kpn {
+
+template<typename T> class Channel; // forward declaration for acquire_balanced
+
+// ── SharedResource ────────────────────────────────────────────────────────────
+//
+// Wraps an exclusive resource (e.g. an ONNX session, a CUDA stream) and
+// arbitrates concurrent access using a priority-based waiter queue.
+//
+// When multiple nodes compete, the one with the highest priority score wins
+// the next slot. Priority is re-evaluated at release time so it reflects the
+// current queue state, not the state when the node first started waiting.
+//
+// Starvation prevention: each waiter's effective score grows with elapsed wait
+// time (aging_per_second), ensuring a low-priority node eventually gets served.
+//
+// Usage:
+//   SharedResource<OrtSession> res(session_args...);
+//
+//   // inside a node functor —
+//   auto guard = res.acquire_balanced(in_channel, out_channel);
+//   guard->Run(...);   // guard releases automatically on scope exit
+
+template<typename T>
+class SharedResource : public IResourceProbe {
+public:
+    // ── RAII guard ────────────────────────────────────────────────────────────
+
+    class Guard {
+        SharedResource* owner_;
+        explicit Guard(SharedResource* o) : owner_(o) {}
+        friend class SharedResource;
+    public:
+        Guard(Guard&& o) noexcept : owner_(std::exchange(o.owner_, nullptr)) {}
+        Guard& operator=(Guard&&) = delete;
+        Guard(const Guard&)       = delete;
+        Guard& operator=(const Guard&) = delete;
+        ~Guard() { if (owner_) owner_->release(); }
+
+        T&  get()          { return owner_->resource_; }
+        T*  operator->()   { return &owner_->resource_; }
+        T&  operator*()    { return owner_->resource_; }
+    };
+
+    // ── Construction ──────────────────────────────────────────────────────────
+
+    template<typename... Args>
+    explicit SharedResource(Args&&... args)
+        : resource_(std::forward<Args>(args)...) {}
+
+    SharedResource(const SharedResource&)            = delete;
+    SharedResource& operator=(const SharedResource&) = delete;
+    SharedResource(SharedResource&&)                 = delete;
+    SharedResource& operator=(SharedResource&&)      = delete;
+
+    // ── Acquire ───────────────────────────────────────────────────────────────
+
+    // Acquire with a callable that returns a priority in [0, 1].
+    // Higher = more urgent. Called at every release to pick the best waiter.
+    template<typename PriorityFn>
+    Guard acquire(PriorityFn&& fn) {
+        std::unique_lock lock(mutex_);
+        if (!held_) {
+            held_ = true;
+            acq_.fetch_add(1, std::memory_order_relaxed);
+            return Guard(this);
+        }
+        Waiter w{std::function<float()>(std::forward<PriorityFn>(fn)), clock_t::now()};
+        waiters_.push_back(&w);
+        update_peak(waiters_.size());
+        current_waiters_.store(waiters_.size(), std::memory_order_relaxed);
+
+        auto t0 = w.wait_start;
+        w.cv.wait(lock, [&w] { return w.ready; });
+
+        int64_t wait_us = std::chrono::duration_cast<std::chrono::microseconds>(
+            clock_t::now() - t0).count();
+        waiters_.erase(std::find(waiters_.begin(), waiters_.end(), &w));
+        current_waiters_.store(waiters_.size(), std::memory_order_relaxed);
+        acq_.fetch_add(1, std::memory_order_relaxed);
+        total_wait_us_.fetch_add(static_cast<uint64_t>(wait_us > 0 ? wait_us : 0),
+                                 std::memory_order_relaxed);
+        return Guard(this);
+    }
+
+    // Acquire with no priority (all waiters treated equally, order is fair-ish).
+    Guard acquire() {
+        return acquire([] { return 0.5f; });
+    }
+
+    // Acquire with priority derived from channel fill fractions:
+    //   score = input_fill × output_headroom
+    // A node with a full input queue and empty output queue has the highest
+    // urgency — it has work to do and nowhere to stall downstream.
+    template<typename In, typename Out>
+    Guard acquire_balanced(const Channel<In>& in_ch, const Channel<Out>& out_ch) {
+        return acquire([&in_ch, &out_ch] {
+            float in_fill  = in_ch.capacity()  ? float(in_ch.size())  / in_ch.capacity()  : 0.5f;
+            float out_head = out_ch.capacity() ? 1.0f - float(out_ch.size()) / out_ch.capacity() : 0.5f;
+            return in_fill * out_head;
+        });
+    }
+
+    // ── IResourceProbe ────────────────────────────────────────────────────────
+
+    ResourceSnapshot snapshot(const std::string& name) const override {
+        std::lock_guard lock(mutex_);
+        uint64_t a   = acq_.load(std::memory_order_relaxed);
+        uint64_t w   = total_wait_us_.load(std::memory_order_relaxed);
+        return {
+            name,
+            a,
+            a > 0 ? double(w) / a / 1000.0 : 0.0,
+            peak_waiters_.load(std::memory_order_relaxed),
+            current_waiters_.load(std::memory_order_relaxed),
+            held_,
+        };
+    }
+
+private:
+    void release() {
+        std::unique_lock lock(mutex_);
+        if (waiters_.empty()) {
+            held_ = false;
+            return;
+        }
+        // Re-evaluate every waiter's current priority and apply aging bonus.
+        auto now = clock_t::now();
+        Waiter* best       = nullptr;
+        float   best_score = -1.0f;
+        for (Waiter* w : waiters_) {
+            float age_s = std::chrono::duration<float>(now - w->wait_start).count();
+            float score = w->priority_fn() + age_s * kAgingPerSecond;
+            if (score > best_score) { best_score = score; best = w; }
+        }
+        best->ready = true;
+        best->cv.notify_one();
+        // held_ stays true — ownership transfers to the woken waiter.
+    }
+
+    void update_peak(std::size_t n) {
+        uint64_t prev = peak_waiters_.load(std::memory_order_relaxed);
+        while (n > prev &&
+               !peak_waiters_.compare_exchange_weak(prev, n,
+                   std::memory_order_relaxed, std::memory_order_relaxed))
+            ;
+    }
+
+    struct Waiter {
+        std::function<float()>  priority_fn;
+        clock_t::time_point     wait_start;
+        std::condition_variable cv;
+        bool                    ready{false};
+
+        Waiter(std::function<float()> fn, clock_t::time_point t)
+            : priority_fn(std::move(fn)), wait_start(t) {}
+    };
+
+    static constexpr float kAgingPerSecond = 0.05f;
+
+    T                        resource_;
+    bool                     held_{false};
+    mutable std::mutex       mutex_;
+    std::vector<Waiter*>     waiters_;
+    std::atomic<uint64_t>    acq_{0};
+    std::atomic<uint64_t>    total_wait_us_{0};
+    std::atomic<uint64_t>    peak_waiters_{0};
+    std::atomic<uint64_t>    current_waiters_{0};
+};
+
+// ── Factory ───────────────────────────────────────────────────────────────────
+
+template<typename T, typename... Args>
+SharedResource<T> make_shared_resource(Args&&... args) {
+    return SharedResource<T>(std::forward<Args>(args)...);
+}
+
+} // namespace kpn

include/kpn/static_network.hpp

@@ -2,7 +2,7 @@
 #include "channel.hpp"
 #include "diagnostics.hpp"
 #include "fanout.hpp"
-#include "node.hpp"
+#include "inode.hpp"
 #include "port.hpp"
 #include "tmp/fanout_groups.hpp"
 #include "tmp/topo_sort.hpp"
@@ -13,7 +13,9 @@
 #endif
 
 #include <iostream>
+#include <map>
 #include <string>
+#include <thread>
 #include <tuple>
 #include <type_traits>
 #include <utility>
@@ -74,6 +76,10 @@ std::string node_display_name() {
         return std::string(lbl);
     } else if constexpr (requires { NodeT::is_fanout_node; NodeT::unique_tag; }) {
         return "fanout[" + std::to_string(NodeT::unique_tag) + "]";
+    } else if constexpr (requires { NodeT::is_router_node; NodeT::unique_tag; }) {
+        return "router[" + std::to_string(NodeT::unique_tag) + "]";
+    } else if constexpr (requires { NodeT::is_filter_node; NodeT::unique_tag; }) {
+        return "filter[" + std::to_string(NodeT::unique_tag) + "]";
     } else if constexpr (requires { NodeT::unique_tag; }) {
         return "node[" + std::to_string(NodeT::unique_tag) + "]";
     } else {
@@ -106,21 +112,35 @@ public:
     void start() override {
         stop_flag_ = false;
         start_time_ = clock_t::now();
+        if (event_handler_) {
+            for (std::size_t i = 0; i < user_nodes_topo_.size(); ++i) {
+                auto* node = user_nodes_topo_[i];
+                const auto& n = user_node_names_[i];
+                node->set_network_overflow_callback(
+                    [this, n](auto ts) { event_handler_(n, NodeEvent::Overflow, ts); });
+                node->set_network_closed_callback(
+                    [this, n](auto ts) { event_handler_(n, NodeEvent::Closed, ts); });
+            }
+        }
         for (auto* n : user_nodes_topo_)  n->start();
         for (auto* n : fanout_nodes_ptr_) n->start();
 #ifdef KPN_WEB_DEBUG
-        web_server_ = std::make_unique<web_debug::WebDebugServer>(
-            web_debug_port_,
-            [this]() {
-                auto s = collect_snapshots();
-                return web_debug::to_json(s.nodes, s.channels, s.elapsed_s);
-            });
-        web_server_->start();
-        std::cerr << "[kpn] web debug UI: http://localhost:" << web_debug_port_ << "\n";
+        if (web_server_enabled_) {
+            web_server_ = std::make_unique<web_debug::WebDebugServer>(
+                web_debug_port_,
+                [this]() {
+                    auto s = collect_snapshots();
+                    return web_debug::to_json(s.nodes, s.channels, s.resources, s.elapsed_s, s.pools);
+                });
+            web_server_->start();
+            std::cerr << "[kpn] web debug UI: http://localhost:" << web_debug_port_ << "\n";
+        }
 #endif
     }
 
-    void stop() override {
+    void stop() override { halt(); }
+
+    void halt() override {
         stop_flag_ = true;
 #ifdef KPN_WEB_DEBUG
         if (web_server_) web_server_->stop();
@@ -131,6 +151,22 @@ public:
             (*it)->stop();
     }
 
+    // shutdown(): graceful drain in topological order (sources first).
+    // Stops source nodes, polls channels until empty, then stops each downstream layer.
+    void shutdown() override {
+        stop_flag_ = true;
+#ifdef KPN_WEB_DEBUG
+        if (web_server_) web_server_->stop();
+#endif
+        // user_nodes_topo_ is already in sources-first order.
+        // Stop each node and drain its output channels before moving on.
+        for (auto* n : user_nodes_topo_) {
+            n->stop();
+            drain_all_channels();
+        }
+        for (auto* n : fanout_nodes_ptr_) n->stop();
+    }
+
     bool running() const override { return !stop_flag_; }
 
     void set_name(std::string name) override { name_ = std::move(name); }
@@ -139,10 +175,36 @@ public:
         return {n, 0, 0, 0, 0, 0, 0, 0};
     }
 
+    using EventHandler =
+        std::function<void(std::string_view node_name, NodeEvent,
+                           std::chrono::steady_clock::time_point)>;
+
+    void set_event_handler(EventHandler h) { event_handler_ = std::move(h); }
+
 #ifdef KPN_WEB_DEBUG
     void set_web_debug_port(uint16_t port) { web_debug_port_ = port; }
+    // Called by DebugHub::register_network() so the hub owns the debug server.
+    void disable_web_server() { web_server_enabled_ = false; }
 #endif
 
+    // Returns a snapshot of this network's nodes and channels for the DebugHub.
+    NetworkSnapshot network_snapshot() const {
+        auto s = collect_snapshots();
+        return {"", std::move(s.nodes), std::move(s.channels), s.elapsed_s};
+    }
+
+    // Register a shared resource so it appears in diagnostics and the debug UI.
+    // The probe must outlive this network (typically the resource is on the same stack).
+    void register_resource(const std::string& name, IResourceProbe* probe) {
+        resource_probes_.emplace_back(name, probe);
+    }
+
+    // Register a thread pool so it appears in diagnostics and the debug UI.
+    // The probe must outlive this network (typically the pool is on the same stack/shared_ptr).
+    void register_pool(const std::string& name, IPoolProbe* probe) {
+        pool_probes_.emplace_back(name, probe);
+    }
+
     // Print diagnostics using compile-time node labels
     void print_diagnostics(std::ostream& os = std::cerr) const {
         os << "\n┌─ KPN++ StaticNetwork diagnostics ─────────────────────────────\n";
@@ -159,9 +221,11 @@ public:
 
 private:
     struct Snapshots {
-        std::vector<NodeSnapshot>    nodes;
-        std::vector<ChannelSnapshot> channels;
-        double                       elapsed_s;
+        std::vector<NodeSnapshot>     nodes;
+        std::vector<ChannelSnapshot>  channels;
+        std::vector<ResourceSnapshot> resources;
+        std::vector<PoolSnapshot>     pools;
+        double                        elapsed_s;
     };
 
     Snapshots collect_snapshots() const {
@@ -178,7 +242,27 @@ private:
         for (auto& probe : channel_probes_)
             channels.push_back(probe->snapshot());
 
-        return {std::move(nodes), std::move(channels), elapsed_s};
+        std::vector<ResourceSnapshot> resources;
+        for (auto& [name, probe] : resource_probes_)
+            resources.push_back(probe->snapshot(name));
+
+        std::vector<PoolSnapshot> pools;
+        for (auto& [name, probe] : pool_probes_)
+            pools.push_back(probe->snapshot(name));
+
+        return {std::move(nodes), std::move(channels), std::move(resources), std::move(pools), elapsed_s};
+    }
+
+    void drain_all_channels() const {
+        bool any_full = true;
+        while (any_full) {
+            any_full = false;
+            for (auto& probe : channel_probes_) {
+                if (probe->snapshot().current_fill > 0) { any_full = true; break; }
+            }
+            if (any_full)
+                std::this_thread::sleep_for(std::chrono::milliseconds(1));
+        }
     }
 
     std::string                                         name_;
@@ -189,9 +273,13 @@ private:
     std::vector<std::string>                            user_node_names_;
     std::vector<std::string>                            fanout_node_names_;
     std::vector<std::unique_ptr<IChannelProbe>>         channel_probes_;
+    std::vector<std::pair<std::string, IResourceProbe*>> resource_probes_;
+    std::vector<std::pair<std::string, IPoolProbe*>>     pool_probes_;
+    EventHandler                                        event_handler_;
     clock_t::time_point                                 start_time_;
 #ifdef KPN_WEB_DEBUG
     uint16_t                                            web_debug_port_{9090};
+    bool                                                web_server_enabled_{true};
     std::unique_ptr<web_debug::WebDebugServer>          web_server_;
 #endif
 };
@@ -229,12 +317,16 @@ auto make_network(Edges&&... edges) {
         auto* s = static_cast<INode*>(&e.src);
         auto* d = static_cast<INode*>(&e.dst);
         if (std::find(user_node_ptrs.begin(), user_node_ptrs.end(), s) == user_node_ptrs.end()) {
+            auto sname = node_display_name<SrcT>();
             user_node_ptrs.push_back(s);
-            user_node_names.push_back(node_display_name<SrcT>());
+            user_node_names.push_back(sname);
+            s->set_name(sname);
         }
         if (std::find(user_node_ptrs.begin(), user_node_ptrs.end(), d) == user_node_ptrs.end()) {
+            auto dname = node_display_name<DstT>();
             user_node_ptrs.push_back(d);
-            user_node_names.push_back(node_display_name<DstT>());
+            user_node_names.push_back(dname);
+            d->set_name(dname);
         }
     };
     (collect(edges), ...);
@@ -317,8 +409,10 @@ auto make_network(Edges&&... edges) {
     std::apply([&](auto&... fn) {
         ([&](auto& node) {
             using NodeT = std::decay_t<decltype(node)>;
+            auto fname = node_name.template operator()<NodeT>();
+            static_cast<INode&>(node).set_name(fname);
             fanout_ptrs.push_back(static_cast<INode*>(&node));
-            fanout_node_names.push_back(node_name.template operator()<NodeT>());
+            fanout_node_names.push_back(fname);
         }(fn), ...);
     }, *fanout_storage);
 

include/kpn/traits.hpp

@@ -83,4 +83,18 @@ template<typename F>
 inline constexpr std::size_t output_count_v =
     std::tuple_size_v<normalised_return_t<return_t<F>>>;
 
+// ── repeat_tuple: std::tuple<T, T, ..., T> with N repetitions ────────────────
+
+template<typename T, std::size_t N, typename Seq = std::make_index_sequence<N>>
+struct repeat_tuple;
+
+template<typename T, std::size_t N, std::size_t... Is>
+struct repeat_tuple<T, N, std::index_sequence<Is...>> {
+    template<std::size_t> using always_T = T;
+    using type = std::tuple<always_T<Is>...>;
+};
+
+template<typename T, std::size_t N>
+using repeat_tuple_t = typename repeat_tuple<T, N>::type;
+
 } // namespace kpn

include/kpn/variant_node.hpp

@@ -239,11 +239,13 @@ private:
 };
 
 // ── PythonConverter ───────────────────────────────────────────────────────────
+// Specialise for each type you want to use across a PyNetwork.
+// Required members: to_python(const T&), from_python(nb::object).
+// Optional member:  static constexpr const char* type_name = "friendly_name";
+// Built-in specialisations for int/float/double/bool/std::string are provided
+// automatically when you include <kpn/python/auto_bind.hpp>.
 
 template<typename T>
-struct PythonConverter {
-    static_assert(sizeof(T) == 0,
-        "PythonConverter<T> must be specialised for each type used in a PyNetwork");
-};
+struct PythonConverter {};
 
 } // namespace kpn

include/kpn/web_debug.hpp

@@ -46,7 +46,9 @@ static std::pair<std::string,std::string> parse_edge_name(const std::string& nam
 
 static std::string to_json(const std::vector<NodeSnapshot>& nodes,
                             const std::vector<ChannelSnapshot>& channels,
-                            double elapsed_s = 0.0) {
+                            const std::vector<ResourceSnapshot>& resources = {},
+                            double elapsed_s = 0.0,
+                            const std::vector<PoolSnapshot>& pools = {}) {
     std::ostringstream o;
     o << std::fixed;
     o.precision(2);
@@ -84,6 +86,32 @@ static std::string to_json(const std::vector<NodeSnapshot>& nodes,
           << ",\"bw_mbs\":"      << c.bandwidth_mbs(elapsed_s)
           << "}";
     }
+    o << "],\"resources\":[";
+    for (std::size_t i = 0; i < resources.size(); ++i) {
+        const auto& r = resources[i];
+        if (i) o << ',';
+        o << "{\"name\":\""          << escape_json(r.name) << "\""
+          << ",\"acquisitions\":"    << r.acquisitions
+          << ",\"avg_wait_ms\":"     << r.avg_wait_ms
+          << ",\"peak_waiters\":"    << r.peak_waiters
+          << ",\"current_waiters\":" << r.current_waiters
+          << ",\"held\":"            << (r.held ? "true" : "false")
+          << "}";
+    }
+    o << "],\"pools\":[";
+    for (std::size_t i = 0; i < pools.size(); ++i) {
+        const auto& p = pools[i];
+        if (i) o << ',';
+        double in_rate  = elapsed_s > 0.0 ? p.tasks_submitted / elapsed_s : 0.0;
+        double out_rate = elapsed_s > 0.0 ? p.tasks_completed / elapsed_s : 0.0;
+        o << "{\"name\":\""        << escape_json(p.name) << "\""
+          << ",\"thread_count\":"  << p.thread_count
+          << ",\"queue_depth\":"   << p.queue_depth
+          << ",\"active_count\":"  << p.active_count
+          << ",\"in_rate\":"       << in_rate
+          << ",\"out_rate\":"      << out_rate
+          << "}";
+    }
     o << "]}";
     return o.str();
 }
@@ -112,12 +140,24 @@ static const char* HTML = R"html(<!DOCTYPE html>
     border-radius: 4px; padding: 8px 12px; font-size: 11px; pointer-events: none;
     display: none; white-space: pre; line-height: 1.6;
   }
+  #resources {
+    position: absolute; bottom: 12px; right: 12px;
+    background: #16213e; border: 1px solid #0f3460; border-radius: 4px;
+    padding: 8px 12px; font-size: 10px; min-width: 220px;
+    display: none;
+  }
+  #resources h2 { margin: 0 0 6px; font-size: 11px; color: #e94560; }
+  .res-row { display: flex; justify-content: space-between; gap: 12px; margin-top: 3px; }
+  .res-name { color: #eee; }
+  .res-held-y { color: #e94560; }
+  .res-held-n { color: #4CAF50; }
 </style>
 </head>
 <body>
 <div id="header"><h1>KPN++ Web Debug</h1><span id="status">connecting...</span></div>
 <svg id="graph"></svg>
 <div id="tooltip"></div>
+<div id="resources"><h2>Shared Resources</h2><div id="res-list"></div></div>
 <script src="https://d3js.org/d3.v7.min.js"></script>
 <script>
 const nodeRadius = 30;
@@ -139,7 +179,7 @@ const defs = svg.append('defs');
 const g = svg.append('g');
 svg.call(d3.zoom().on('zoom', e => g.attr('transform', e.transform)));
 
-let sim, linkSel, nodeSel, labelSel, edgeLabelSel;
+let sim, linkSel, nodeSel, labelSel, edgeLabelSel, edgeLabelBwSel;
 let nodes = [], links = [];
 
 function edgeColor(fill_pct) {
@@ -186,6 +226,10 @@ function init(data) {
     .attr('class', 'link-label')
     .text(d => `${d.fill_pct.toFixed(0)}%`);
 
+  edgeLabelBwSel = g.append('g').selectAll('text').data(links).join('text')
+    .attr('class', 'link-label')
+    .text(d => `${d.bw_mbs.toFixed(1)} MB/s`);
+
   // Nodes
   const nodeG = g.append('g').selectAll('g').data(nodes).join('g')
     .attr('class', 'node')
@@ -200,6 +244,7 @@ function init(data) {
     .text(d => `${d.ema_exec_ms.toFixed(1)}ms ${d.fps.toFixed(1)}fps`);
 
   nodeSel = nodeG;
+  renderResources(data.resources);
 
   // Tooltips
   const tip = d3.select('#tooltip');
@@ -224,7 +269,25 @@ function init(data) {
   }).on('mouseleave', () => tip.style('display','none'));
 }
 
+function renderResources(resources) {
+  const panel = document.getElementById('resources');
+  const list  = document.getElementById('res-list');
+  if (!resources || resources.length === 0) { panel.style.display = 'none'; return; }
+  panel.style.display = 'block';
+  list.innerHTML = resources.map(r => {
+    const heldCls = r.held ? 'res-held-y' : 'res-held-n';
+    const heldTxt = r.held ? 'HELD' : 'free';
+    return `<div class="res-row">
+      <span class="res-name">${r.name}</span>
+      <span class="${heldCls}">${heldTxt}</span>
+      <span>wait ${r.avg_wait_ms.toFixed(1)}ms</span>
+      <span>waiters ${r.current_waiters}/${r.peak_waiters}</span>
+    </div>`;
+  }).join('');
+}
+
 function update(data) {
+  renderResources(data.resources);
   // Update node stats in-place (preserve simulation x/y positions)
   const byId = Object.fromEntries(data.nodes.map(n => [n.id, n]));
   nodes.forEach(n => {
@@ -252,6 +315,7 @@ function update(data) {
   linkSel.attr('stroke', d => edgeColor(d.fill_pct))
          .attr('marker-end', d => edgeArrow(d.fill_pct));
   edgeLabelSel.text(d => `${d.fill_pct.toFixed(0)}%`);
+  edgeLabelBwSel.text(d => `${d.bw_mbs.toFixed(1)} MB/s`);
 }
 
 function ticked() {
@@ -278,6 +342,10 @@ function ticked() {
     .attr('x', d => (d.source.x + d.target.x)/2)
     .attr('y', d => (d.source.y + d.target.y)/2 - 6);
 
+  edgeLabelBwSel
+    .attr('x', d => (d.source.x + d.target.x)/2)
+    .attr('y', d => (d.source.y + d.target.y)/2 + 8);
+
   nodeSel.attr('transform', d => `translate(${d.x},${d.y})`);
 }
 
@@ -311,12 +379,14 @@ class WebDebugServer {
 public:
     using SnapshotFn = std::function<std::string()>;
 
-    explicit WebDebugServer(uint16_t port, SnapshotFn fn)
-        : port_(port), snapshot_fn_(std::move(fn)) {}
+    explicit WebDebugServer(uint16_t port, SnapshotFn fn,
+                            const char* custom_html = nullptr)
+        : port_(port), snapshot_fn_(std::move(fn))
+        , html_(custom_html ? custom_html : HTML) {}
 
     void start() {
-        svr_.Get("/", [](const httplib::Request&, httplib::Response& res) {
-            res.set_content(HTML, "text/html");
+        svr_.Get("/", [this](const httplib::Request&, httplib::Response& res) {
+            res.set_content(html_, "text/html");
         });
         svr_.Get("/api/snapshot", [this](const httplib::Request&, httplib::Response& res) {
             res.set_content(snapshot_fn_(), "application/json");
@@ -337,10 +407,11 @@ public:
     WebDebugServer& operator=(const WebDebugServer&) = delete;
 
 private:
-    uint16_t     port_;
-    SnapshotFn   snapshot_fn_;
+    uint16_t        port_;
+    SnapshotFn      snapshot_fn_;
+    const char*     html_;
     httplib::Server svr_;
-    std::thread  thread_;
+    std::thread     thread_;
 };
 
 } // namespace kpn::web_debug

mkdocs.yml

@@ -0,0 +1,50 @@
+site_name: KPN++
+site_description: A C++20 Kahn Process Network library
+repo_url: https://github.com/yourusername/kpn
+repo_name: kpn
+
+theme:
+  name: material
+  palette:
+    - scheme: slate
+      primary: indigo
+      accent: indigo
+  features:
+    - navigation.tabs
+    - navigation.sections
+    - navigation.top
+    - content.code.copy
+    - content.code.annotate
+
+nav:
+  - Home: index.md
+  - Getting Started: getting-started.md
+  - Concepts:
+    - Nodes: nodes.md
+    - Networks: network.md
+    - Channels: channels.md
+  - Error Handling & Events: error-handling.md
+  - Advanced:
+    - Static Networks: static-network.md
+    - Shared Resources: shared-resource.md
+    - Fan-out & Routing: fanout.md
+  - Examples: examples.md
+
+markdown_extensions:
+  - admonition
+  - toc:
+      permalink: true
+  - pymdownx.highlight:
+      anchor_linenums: true
+      line_spans: __span
+      pygments_lang_class: true
+  - pymdownx.inlinehilite
+  - pymdownx.superfences
+  - pymdownx.tabbed:
+      alternate_style: true
+  - pymdownx.snippets:
+      base_path: ['.']
+      check_paths: true
+  - pymdownx.details
+  - attr_list
+  - md_in_html

python/kpn_python.cpp

@@ -1,164 +1,39 @@
 #define KPN_BUILD_PYTHON
-#include <kpn/python/bindings.hpp>
-#include <nanobind/nanobind.h>
-#include <nanobind/stl/shared_ptr.h>
-#include <nanobind/stl/string.h>
-#include <nanobind/stl/vector.h>
+#include <kpn/python/auto_bind.hpp>
 
 #include <iostream>
-#include <map>
 
 namespace nb = nanobind;
 using namespace kpn;
 using namespace kpn::python;
 
-// ── Node functions for the hello-pipeline examples ────────────────────────────
+// ── Demo node functions (hello-pipeline) ──────────────────────────────────────
 
 static int  produce()        { return 42; }
 static int  double_it(int x) { return x * 2; }
 static void print_it(int x)  { std::cout << "result: " << x << '\n'; }
 
-// ── Variant type ──────────────────────────────────────────────────────────────
-// Deduplicated port types across all registered node functions.
-// produce: () → int  |  double_it: int → int  |  print_it: int → void
-// Unique types: int
+// ── Registry ──────────────────────────────────────────────────────────────────
+// Variant type is auto-deduced as std::variant<int> from the port types above.
 
-using KpnVariant  = std::variant<int>;
-using Net         = PyNetwork<KpnVariant>;
-using ProduceNode = VariantNodeWrapper<produce,   KpnVariant>;
-using DoubleItNode= VariantNodeWrapper<double_it,  KpnVariant>;
-using PrintItNode = VariantNodeWrapper<print_it,   KpnVariant>;
+using DemoNodes = NodeRegistry<
+    Entry<produce,   "produce">,
+    Entry<double_it, "double_it">,
+    Entry<print_it,  "print_it">
+>;
 
-// ── Converter helpers for int ─────────────────────────────────────────────────
-
-static nb::object int_to_py(const KpnVariant& v) {
-    return nb::int_(std::get<int>(v));
-}
-
-static KpnVariant int_from_py(nb::object o) {
-    return KpnVariant{ nb::cast<int>(o) };
-}
-
-// ── Type name resolver (extensible) ──────────────────────────────────────────
-
-static std::type_index resolve_type(const std::string& name) {
-    if (name == "int") return std::type_index(typeid(int));
-    throw std::runtime_error("unknown type name '" + name +
-                             "' — only 'int' is registered in this module");
-}
-
-// ── Ensure converters are registered on a Net instance ───────────────────────
-
-static void ensure_converters(Net& net) {
-    net.register_type<int>(
-        [](const int& v) -> nb::object { return nb::int_(v); },
-        [](nb::object o) -> int        { return nb::cast<int>(o); }
-    );
-    net.register_tap_factory<int>();
-}
+// ── Module ────────────────────────────────────────────────────────────────────
 
 NB_MODULE(kpn_python, m) {
     m.doc() = "KPN++ Python bindings — Kahn Process Network library";
 
-    // ── IVariantNode base ─────────────────────────────────────────────────────
-    nb::class_<IVariantNode<KpnVariant>>(m, "INode");
+    // Registers: Network, INode, ProduceNode, DoubleItNode, PrintItNode,
+    //            make_produce(), make_double_it(), make_print_it()
+    bind_network<DemoNodes>(m);
 
-    // ── Concrete C++ node wrappers ─────────────────────────────────────────────
-    // Factories return shared_ptr so Python can pass them to net.add().
-    nb::class_<ProduceNode,  IVariantNode<KpnVariant>>(m, "ProduceNode")
-        .def("__init__", [](ProduceNode* self, std::size_t cap) {
-            new (self) ProduceNode(cap);
-        }, nb::arg("capacity") = 5);
-
-    nb::class_<DoubleItNode, IVariantNode<KpnVariant>>(m, "DoubleItNode")
-        .def("__init__", [](DoubleItNode* self, std::size_t cap) {
-            new (self) DoubleItNode(cap);
-        }, nb::arg("capacity") = 5);
-
-    nb::class_<PrintItNode,  IVariantNode<KpnVariant>>(m, "PrintItNode")
-        .def("__init__", [](PrintItNode* self, std::size_t cap) {
-            new (self) PrintItNode(cap);
-        }, nb::arg("capacity") = 5);
-
-    // Expose factory functions that return shared_ptr — these are what net.add() accepts.
-    m.def("make_produce",   [](std::size_t cap) -> std::shared_ptr<IVariantNode<KpnVariant>> {
-        return std::make_shared<ProduceNode>(cap);
-    }, nb::arg("capacity") = 5);
-    m.def("make_double_it", [](std::size_t cap) -> std::shared_ptr<IVariantNode<KpnVariant>> {
-        return std::make_shared<DoubleItNode>(cap);
-    }, nb::arg("capacity") = 5);
-    m.def("make_print_it",  [](std::size_t cap) -> std::shared_ptr<IVariantNode<KpnVariant>> {
-        return std::make_shared<PrintItNode>(cap);
-    }, nb::arg("capacity") = 5);
-
-    // ── Network ───────────────────────────────────────────────────────────────
-    nb::class_<Net>(m, "Network")
-        .def(nb::init<>())
-
-        // add(name, c++_node) — for pre-constructed C++ nodes
-        .def("add", [](Net& self, std::string name,
-                       std::shared_ptr<IVariantNode<KpnVariant>> node) {
-            self.add(std::move(name), std::move(node));
-        }, nb::arg("name"), nb::arg("node"))
-
-        // add_node(name, callable, inputs=[type_names], outputs=[type_names])
-        // Creates a pure Python processing node.
-        .def("add_node", [](Net& self,
-                            std::string name,
-                            nb::object callable,
-                            std::vector<std::string> in_names,
-                            std::vector<std::string> out_names,
-                            std::size_t capacity)
-        {
-            std::vector<std::type_index> in_types, out_types;
-            for (auto& s : in_names)  in_types.push_back(resolve_type(s));
-            for (auto& s : out_names) out_types.push_back(resolve_type(s));
-
-            std::map<std::type_index, std::function<nb::object(const KpnVariant&)>> to_py;
-            std::map<std::type_index, std::function<KpnVariant(nb::object)>>        from_py;
-            std::map<std::type_index, PyNode<KpnVariant>::ChannelFactory>            ch_factories;
-
-            auto int_idx = std::type_index(typeid(int));
-            to_py[int_idx]      = int_to_py;
-            from_py[int_idx]    = int_from_py;
-            ch_factories[int_idx] = [](std::size_t cap) -> std::shared_ptr<IVariantChannel<KpnVariant>> {
-                auto ch = std::make_shared<Channel<int>>(cap);
-                return std::make_shared<VariantChannel<int, KpnVariant>>(std::move(ch));
-            };
-
-            auto node = std::make_shared<PyNode<KpnVariant>>(
-                std::move(callable),
-                std::move(in_types),
-                std::move(out_types),
-                std::move(to_py),
-                std::move(from_py),
-                std::move(ch_factories),
-                capacity
-            );
-            self.add(std::move(name), std::move(node));
-        },
-        nb::arg("name"), nb::arg("callable"),
-        nb::arg("inputs")   = std::vector<std::string>{},
-        nb::arg("outputs")  = std::vector<std::string>{},
-        nb::arg("capacity") = 5)
-
-        .def("connect", &Net::connect,
-             nb::arg("src"), nb::arg("out_idx"),
-             nb::arg("dst"), nb::arg("in_idx"))
-        .def("build",  &Net::build)
-        .def("start",  &Net::start)
-        .def("stop",   &Net::stop)
-
-        // read(node, out_idx=0) — blocking pop from a C++ node's output port into Python
-        .def("read", [](Net& self, const std::string& node, std::size_t out_idx) {
-            ensure_converters(self);
-            return self.read(node, out_idx);
-        }, nb::arg("node"), nb::arg("out_idx") = 0)
-
-        // write(node, in_idx, value) — push a Python value into a node's input port
-        .def("write", [](Net& self, const std::string& node,
-                         std::size_t in_idx, nb::object value) {
-            ensure_converters(self);
-            self.write(node, in_idx, std::move(value));
-        }, nb::arg("node"), nb::arg("in_idx"), nb::arg("value"));
+    // Also expose raw functions for direct testing (no network needed):
+    //   kpn.produce()     → 42
+    //   kpn.double_it(5)  → 10
+    //   kpn.print_it(84)  → prints "result: 84", returns None
+    bind_debug<DemoNodes>(m);
 }

scripts/render_readme.py

@@ -0,0 +1,121 @@
+#!/usr/bin/env python3
+"""
+Render README.md from README.md.in by expanding @snippet directives.
+
+Directive format (in README.md.in):
+  <!-- @snippet path/to/file.cpp snippet_name -->
+
+Snippet tags (in C++ source files):
+  // [snippet: snippet_name]
+  ...content...
+  // [/snippet: snippet_name]
+
+In Python files use # instead of //.
+"""
+
+import re
+import sys
+import textwrap
+from pathlib import Path
+
+ROOT = Path(__file__).parent.parent
+TEMPLATE = ROOT / "README.md.in"
+OUTPUT = ROOT / "README.md"
+
+DIRECTIVE_RE = re.compile(r'<!--\s*@snippet\s+(\S+)\s+(\S+)\s*-->')
+
+LANG_MAP = {'.cpp': 'cpp', '.hpp': 'cpp', '.h': 'cpp', '.py': 'python', '.c': 'c'}
+
+
+def comment_prefix(path: Path) -> str:
+    return '#' if path.suffix == '.py' else '//'
+
+
+def extract_snippet(file_path: Path, name: str) -> str:
+    prefix = comment_prefix(file_path)
+    open_tag  = f'{prefix} [snippet: {name}]'
+    close_tag = f'{prefix} [/snippet: {name}]'
+
+    text = file_path.read_text()
+    lines = text.splitlines()
+
+    in_snippet = False
+    found = False
+    result = []
+
+    for line in lines:
+        stripped = line.strip()
+        if stripped == open_tag:
+            if in_snippet:
+                raise ValueError(f"Nested open tag for '{name}' in {file_path}")
+            in_snippet = True
+            found = True
+            continue
+        if stripped == close_tag:
+            if not in_snippet:
+                raise ValueError(f"Unexpected close tag for '{name}' in {file_path}")
+            in_snippet = False
+            continue
+        if in_snippet:
+            result.append(line)
+
+    if not found:
+        raise ValueError(f"Snippet '{name}' not found in {file_path}")
+    if in_snippet:
+        raise ValueError(f"Unclosed snippet '{name}' in {file_path}")
+
+    # Strip leading/trailing blank lines
+    while result and not result[0].strip():
+        result.pop(0)
+    while result and not result[-1].strip():
+        result.pop()
+
+    # Dedent common leading whitespace
+    content = textwrap.dedent('\n'.join(result))
+    return content
+
+
+def lang_for(file_path: Path) -> str:
+    return LANG_MAP.get(file_path.suffix, '')
+
+
+def render(template: str) -> str:
+    errors = []
+
+    def replace(m: re.Match) -> str:
+        rel_path = m.group(1)
+        name = m.group(2)
+        file_path = ROOT / rel_path
+        if not file_path.exists():
+            errors.append(f"File not found: {file_path}")
+            return m.group(0)
+        try:
+            content = extract_snippet(file_path, name)
+        except ValueError as e:
+            errors.append(str(e))
+            return m.group(0)
+        lang = lang_for(file_path)
+        return f'```{lang}\n{content}\n```'
+
+    result = DIRECTIVE_RE.sub(replace, template)
+
+    if errors:
+        for err in errors:
+            print(f"ERROR: {err}", file=sys.stderr)
+        sys.exit(1)
+
+    return result
+
+
+def main() -> None:
+    if not TEMPLATE.exists():
+        print(f"Error: {TEMPLATE} not found", file=sys.stderr)
+        sys.exit(1)
+
+    rendered = render(TEMPLATE.read_text())
+    OUTPUT.write_text(rendered)
+    print(f"Rendered {TEMPLATE.name} → {OUTPUT.name}")
+
+
+if __name__ == '__main__':
+    main()

src/network.cpp

@@ -1,4 +1,4 @@
-// network.cpp — orchestrator/watchdog implementation details.
+// network.cpp — orchestrator/watchdog implementation details. (CI: pipeline trigger)
 // Most of the Network class is header-only (template-heavy).
 // Non-template implementation lives here once the watchdog grows
 // beyond the stub in network.hpp.

tests/CMakeLists.txt

@@ -32,6 +32,9 @@ add_executable(kpn_tests
     test_node.cpp
     test_network.cpp
     test_static_network.cpp
+    test_shared_resource.cpp
+    test_pool_node.cpp
+    test_scheduler.cpp
 )
 
 target_link_libraries(kpn_tests PRIVATE

tests/test_channel.cpp

@@ -1,6 +1,8 @@
 #include <catch2/catch_test_macros.hpp>
+#include <catch2/catch_approx.hpp>
 #include <kpn/channel.hpp>
 #include <thread>
+#include <vector>
 
 using namespace kpn;
 
@@ -70,13 +72,21 @@ TEST_CASE("push to disabled channel is silently dropped", "[channel]") {
     REQUIRE(ch.size() == 0);
 }
 
-TEST_CASE("disable clears existing queue contents", "[channel]") {
+TEST_CASE("disable stops accepting and unblocks pop", "[channel]") {
+    // With the SPSC lock-free ring, disable() does not drain the ring immediately;
+    // items are freed when the Channel is destroyed. What it must do is:
+    //  1. reject further pushes (drops silently)
+    //  2. unblock any waiting pop() (throws ChannelClosedError)
     Channel<int> ch(5);
     ch.push(1);
     ch.push(2);
     REQUIRE(ch.size() == 2);
     ch.disable();
-    REQUIRE(ch.size() == 0);
+    // Further pushes are dropped
+    ch.push(3);
+    REQUIRE(ch.size() <= 2);
+    // pop() must throw even though items remain in the ring
+    REQUIRE_THROWS_AS(ch.pop(), ChannelClosedError);
 }
 
 TEST_CASE("enable re-accepts pushes after disable", "[channel]") {
@@ -97,3 +107,71 @@ TEST_CASE("large type stored as shared_ptr — no copy on pop", "[channel]") {
     auto out = ch.pop();
     REQUIRE(out.tag == 123);
 }
+
+// ── ChannelDataSize / bytes_pushed tests ─────────────────────────────────────
+
+TEST_CASE("bytes_pushed uses sizeof(T) by default for POD type", "[channel][bandwidth]") {
+    Channel<int> ch(10);
+    ch.push(1);
+    ch.push(2);
+    ch.push(3);
+    ch.pop(); ch.pop(); ch.pop();
+
+    auto snap = ch.snapshot("test");
+    REQUIRE(snap.pushes == 3);
+    REQUIRE(snap.bytes_pushed == 3 * sizeof(int));
+}
+
+TEST_CASE("bytes_pushed uses sizeof(T) by default for large struct", "[channel][bandwidth]") {
+    struct Blob { char data[256]; };
+    Channel<Blob> ch(10);
+    ch.push(Blob{});
+    ch.push(Blob{});
+
+    auto snap = ch.snapshot("test");
+    REQUIRE(snap.bytes_pushed == 2 * sizeof(Blob));
+}
+
+// A fake heap-owning type whose logical payload size differs from sizeof.
+struct FakeFrame {
+    std::vector<uint8_t> pixels;
+};
+
+// Specialise ChannelDataSize so the channel counts actual pixel bytes.
+template<>
+struct kpn::ChannelDataSize<FakeFrame> {
+    static std::size_t bytes(const FakeFrame& f) { return f.pixels.size(); }
+};
+
+TEST_CASE("bytes_pushed uses ChannelDataSize specialisation for heap-owning type", "[channel][bandwidth]") {
+    Channel<FakeFrame> ch(10);
+    ch.push(FakeFrame{std::vector<uint8_t>(1000)});
+    ch.push(FakeFrame{std::vector<uint8_t>(2000)});
+
+    auto snap = ch.snapshot("test");
+    REQUIRE(snap.pushes == 2);
+    REQUIRE(snap.bytes_pushed == 3000);
+    // item_bytes is still sizeof(FakeFrame) — the struct header
+    REQUIRE(snap.item_bytes == sizeof(FakeFrame));
+}
+
+TEST_CASE("bandwidth_mbs is non-zero and correct for heap-owning type", "[channel][bandwidth]") {
+    Channel<FakeFrame> ch(10);
+    // Push 10 frames of 1 MB each
+    for (int i = 0; i < 10; ++i)
+        ch.push(FakeFrame{std::vector<uint8_t>(1'000'000)});
+
+    auto snap = ch.snapshot("test");
+    // 10 MB over 1 second => 10.0 MB/s
+    REQUIRE(snap.bandwidth_mbs(1.0) == Catch::Approx(10.0).epsilon(1e-6));
+    // Without the fix (using sizeof), this would have been ~40 bytes/s ≈ 0.00004 MB/s.
+    REQUIRE(snap.bandwidth_mbs(1.0) > 1.0);
+}
+
+TEST_CASE("bandwidth_mbs returns 0 when elapsed_s is zero or negative", "[channel][bandwidth]") {
+    Channel<int> ch(5);
+    ch.push(42);
+    auto snap = ch.snapshot("test");
+    REQUIRE(snap.bandwidth_mbs(0.0) == 0.0);
+    REQUIRE(snap.bandwidth_mbs(-1.0) == 0.0);
+}

tests/test_node.cpp

@@ -2,6 +2,7 @@
 #include <kpn/node.hpp>
 #include <chrono>
 #include <thread>
+#include <utility>
 
 using namespace kpn;
 
@@ -47,3 +48,104 @@ TEST_CASE("node stop unblocks cleanly", "[node]") {
     node.stop();
     REQUIRE_FALSE(node.running());
 }
+
+// ── Error handler tests ───────────────────────────────────────────────────────
+
+namespace {
+
+struct SometimesThrower {
+    bool throw_next = true;
+    int operator()(int x) {
+        if (std::exchange(throw_next, false))
+            throw std::runtime_error("deliberate skip");
+        return x * 2;
+    }
+};
+
+struct AlwaysThrows {
+    int operator()(int) { throw std::runtime_error("always throws"); return 0; }
+};
+
+static int always_throws_fn(int) {
+    throw std::runtime_error("nttp always throws");
+    return 0;
+}
+
+} // namespace
+
+TEST_CASE("node stops cleanly on exception with no error handler", "[node][error_handler]") {
+    AlwaysThrows obj;
+    auto node = make_node(obj);
+
+    node.start();
+    node.input_channel<0>().push(1);
+    std::this_thread::sleep_for(std::chrono::milliseconds(50));
+    REQUIRE_FALSE(node.running());
+}
+
+TEST_CASE("node continues when error handler returns true", "[node][error_handler]") {
+    SometimesThrower obj;
+    auto node = make_node(obj);
+
+    Channel<int> out_ch(5);
+    node.set_output_channel<0>(&out_ch);
+
+    bool handler_called = false;
+    node.set_error_handler([&](std::string_view, std::exception_ptr) {
+        handler_called = true;
+        return true;
+    });
+
+    node.start();
+    node.input_channel<0>().push(0);   // throws → skipped, no output
+    node.input_channel<0>().push(21);  // succeeds → 42
+    int result = out_ch.pop();         // blocking — waits for the second item
+    node.stop();
+
+    REQUIRE(handler_called);
+    REQUIRE(result == 42);
+}
+
+TEST_CASE("node stops when error handler returns false", "[node][error_handler]") {
+    AlwaysThrows obj;
+    auto node = make_node(obj);
+
+    node.set_error_handler([](std::string_view, std::exception_ptr) { return false; });
+
+    node.start();
+    node.input_channel<0>().push(1);
+    std::this_thread::sleep_for(std::chrono::milliseconds(50));
+    REQUIRE_FALSE(node.running());
+}
+
+TEST_CASE("error handler receives node name and exception", "[node][error_handler]") {
+    AlwaysThrows obj;
+    auto node = make_node(obj);
+    node.set_name("test_node");
+
+    std::string captured_name;
+    std::string captured_msg;
+    node.set_error_handler([&](std::string_view name, std::exception_ptr ep) {
+        captured_name = name;
+        try { std::rethrow_exception(ep); }
+        catch (const std::exception& e) { captured_msg = e.what(); }
+        return false;
+    });
+
+    node.start();
+    node.input_channel<0>().push(1);
+    std::this_thread::sleep_for(std::chrono::milliseconds(50));
+    node.stop();
+
+    REQUIRE(captured_name == "test_node");
+    REQUIRE(captured_msg == "always throws");
+}
+
+TEST_CASE("Node<NTTP> stops cleanly on exception with no error handler", "[node][error_handler]") {
+    auto node = make_node<always_throws_fn>();
+
+    node.start();
+    node.input_channel<0>().push(1);
+    std::this_thread::sleep_for(std::chrono::milliseconds(50));
+    REQUIRE_FALSE(node.running());
+}

tests/test_pool_node.cpp

@@ -0,0 +1,464 @@
+#include <catch2/catch_test_macros.hpp>
+#include <kpn/scheduler.hpp>
+#include <kpn/pool_node.hpp>
+#include <kpn/interrupt_node.hpp>
+#include <atomic>
+#include <chrono>
+#include <mutex>
+#include <thread>
+
+using namespace kpn;
+
+static int double_it(int x) { return x * 2; }
+static std::tuple<int, float> split_it(int x) { return {x, float(x) * 0.5f}; }
+static void consume_it(int x) { (void)x; }
+
+// ── ThreadPool ────────────────────────────────────────────────────────────────
+
+TEST_CASE("thread pool starts and stops cleanly", "[scheduler]") {
+    ThreadPool pool(2);
+    pool.start();
+    pool.stop();
+}
+
+TEST_CASE("thread pool executes submitted tasks", "[scheduler]") {
+    ThreadPool pool(2);
+    pool.start();
+
+    std::atomic<int> counter{0};
+    for (int i = 0; i < 10; ++i)
+        pool.submit([&] { counter.fetch_add(1); });
+
+    pool.drain();
+    REQUIRE(counter.load() == 10);
+    pool.stop();
+}
+
+TEST_CASE("thread pool drain waits for all tasks", "[scheduler]") {
+    ThreadPool pool(1);
+    pool.start();
+
+    std::atomic<bool> done{false};
+    pool.submit([&] {
+        std::this_thread::sleep_for(std::chrono::milliseconds(20));
+        done.store(true);
+    });
+    pool.drain();
+    REQUIRE(done.load());
+    pool.stop();
+}
+
+TEST_CASE("thread pool priority: higher priority tasks run first", "[scheduler]") {
+    ThreadPool pool(1);  // single thread so order is deterministic
+    pool.start();
+
+    // Submit a task that blocks the worker, then queue two tasks with
+    // different priorities. When the blocker finishes, the high-priority
+    // task should run before the low-priority one.
+    std::vector<int> order;
+    std::mutex order_mutex;
+
+    std::atomic<bool> blocker_done{false};
+    pool.submit([&] {
+        std::this_thread::sleep_for(std::chrono::milliseconds(30));
+        blocker_done.store(true);
+    }, 0.5f);
+
+    // Wait until blocker is running, then enqueue the two ordered tasks.
+    while (!blocker_done.load()) std::this_thread::sleep_for(std::chrono::milliseconds(1));
+
+    pool.submit([&] { std::lock_guard g(order_mutex); order.push_back(1); }, 0.1f);
+    pool.submit([&] { std::lock_guard g(order_mutex); order.push_back(2); }, 0.9f);
+
+    pool.drain();
+    REQUIRE(order == std::vector<int>{2, 1});
+    pool.stop();
+}
+
+// ── PoolNode ──────────────────────────────────────────────────────────────────
+
+TEST_CASE("pool node input/output counts", "[pool_node]") {
+    STATIC_REQUIRE(PoolNode<double_it>::input_count  == 1);
+    STATIC_REQUIRE(PoolNode<double_it>::output_count == 1);
+    STATIC_REQUIRE(PoolNode<split_it>::output_count  == 2);
+    STATIC_REQUIRE(PoolNode<consume_it>::output_count == 0);
+}
+
+TEST_CASE("pool node processes items end-to-end", "[pool_node]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    auto node = make_pool_node<double_it>(pool);
+    Channel<int> out_ch(10);
+    node.set_output_channel<0>(&out_ch);
+    node.start();
+
+    node.input_channel<0>().push(21);
+    int result = out_ch.pop();
+
+    node.stop();
+    pool->stop();
+
+    REQUIRE(result == 42);
+}
+
+TEST_CASE("pool node processes multiple items in order", "[pool_node]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    auto node = make_pool_node<double_it>(pool, 20);  // capacity 20
+    Channel<int> out_ch(20);
+    node.set_output_channel<0>(&out_ch);
+    node.start();
+
+    constexpr int N = 10;
+    for (int i = 0; i < N; ++i)
+        node.input_channel<0>().push(i);
+
+    std::vector<int> results;
+    for (int i = 0; i < N; ++i)
+        results.push_back(out_ch.pop());
+
+    node.stop();
+    pool->stop();
+
+    REQUIRE(results.size() == N);
+    for (int i = 0; i < N; ++i)
+        REQUIRE(results[i] == i * 2);
+}
+
+TEST_CASE("pool node stop is clean with no deadlock", "[pool_node]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    auto node = make_pool_node<double_it>(pool);
+    node.start();
+    // Node is idle (no input pushed) — stop must return without deadlock.
+    node.stop();
+    REQUIRE_FALSE(node.running());
+
+    pool->stop();
+}
+
+TEST_CASE("pool node two-stage pipeline produces correct count", "[pool_node]") {
+    auto pool = std::make_shared<ThreadPool>(4);
+    pool->start();
+
+    auto src       = make_pool_node<double_it>(pool);
+    auto transform = make_pool_node<double_it>(pool);
+    Channel<int> out_ch(20);
+
+    // Wire src output → transform input channel, transform output → out_ch.
+    src.set_output_channel<0>(&transform.input_channel<0>());
+    transform.set_output_channel<0>(&out_ch);
+
+    src.start();
+    transform.start();
+
+    constexpr int N = 5;
+    for (int i = 1; i <= N; ++i)
+        src.input_channel<0>().push(i);
+
+    std::vector<int> results;
+    for (int i = 0; i < N; ++i)
+        results.push_back(out_ch.pop());
+
+    // Stop nodes before they (and their channels) go out of scope.
+    src.stop();
+    transform.stop();
+    pool->stop();
+
+    REQUIRE(results.size() == static_cast<std::size_t>(N));
+    for (int i = 0; i < N; ++i)
+        REQUIRE(results[i] == (i + 1) * 4);  // double_it twice
+}
+
+// ── InterruptNode ─────────────────────────────────────────────────────────────
+
+namespace {
+static std::atomic<int> g_interrupt_counter{0};
+static int interrupt_produce() { return g_interrupt_counter.fetch_add(1); }
+} // namespace
+
+TEST_CASE("interrupt node fires on each trigger", "[interrupt_node]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    g_interrupt_counter.store(0);
+    auto node = make_interrupt_node<interrupt_produce>(pool, out<>{});
+    Channel<int> out_ch(20);
+    node.set_output_channel<0>(&out_ch);
+    node.start();
+
+    auto trigger = node.get_trigger();
+    constexpr int N = 5;
+    for (int i = 0; i < N; ++i) trigger();
+
+    std::vector<int> results;
+    for (int i = 0; i < N; ++i)
+        results.push_back(out_ch.pop());
+
+    node.stop();
+    pool->stop();
+
+    REQUIRE(results.size() == static_cast<std::size_t>(N));
+}
+
+TEST_CASE("interrupt node does not fire without trigger", "[interrupt_node]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    g_interrupt_counter.store(0);
+    auto node = make_interrupt_node<interrupt_produce>(pool, out<>{});
+    Channel<int> out_ch(5);
+    node.set_output_channel<0>(&out_ch);
+    node.start();
+
+    std::this_thread::sleep_for(std::chrono::milliseconds(30));
+    // No trigger fired — output channel should be empty.
+    REQUIRE(out_ch.approx_size() == 0);
+
+    node.stop();
+    pool->stop();
+}
+
+TEST_CASE("interrupt node: trigger after stop is ignored", "[interrupt_node]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    g_interrupt_counter.store(0);
+    auto node = make_interrupt_node<interrupt_produce>(pool, out<>{});
+    Channel<int> out_ch(5);
+    node.set_output_channel<0>(&out_ch);
+    node.start();
+
+    auto trigger = node.get_trigger();
+    node.stop();
+    trigger();  // should be a no-op
+
+    std::this_thread::sleep_for(std::chrono::milliseconds(10));
+    REQUIRE(out_ch.approx_size() == 0);
+    pool->stop();
+}
+
+// ── Overflow callback ─────────────────────────────────────────────────────────
+
+TEST_CASE("pool node overflow callback fires on full output channel", "[pool_node][overflow]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    auto node = make_pool_node<double_it>(pool);
+    // Pre-fill a tiny channel so every node push overflows.
+    Channel<int> full_ch(1);
+    full_ch.push(99);
+    node.set_output_channel<0>(&full_ch);
+
+    std::atomic<int> overflow_count{0};
+    node.set_overflow_callback([&](auto) { overflow_count.fetch_add(1); });
+
+    node.start();
+    node.input_channel<0>().push(1);
+    node.input_channel<0>().push(2);
+
+    std::this_thread::sleep_for(std::chrono::milliseconds(50));
+    node.stop();
+    pool->stop();
+
+    REQUIRE(overflow_count.load() > 0);
+}
+
+TEST_CASE("pool node overflow callback is independent per instance", "[pool_node][overflow]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    auto nodeA = make_pool_node<double_it>(pool);
+    auto nodeB = make_pool_node<double_it>(pool);
+
+    std::atomic<int> a_overflows{0}, b_overflows{0};
+    nodeA.set_overflow_callback([&](auto) { a_overflows.fetch_add(1); });
+
+    Channel<int> full_ch(1);
+    full_ch.push(0);
+    nodeA.set_output_channel<0>(&full_ch);
+
+    Channel<int> ok_ch(20);
+    nodeB.set_output_channel<0>(&ok_ch);
+
+    nodeA.start();
+    nodeB.start();
+
+    nodeA.input_channel<0>().push(1);
+    nodeA.input_channel<0>().push(2);
+    nodeB.input_channel<0>().push(10);
+
+    std::this_thread::sleep_for(std::chrono::milliseconds(50));
+    nodeA.stop();
+    nodeB.stop();
+    pool->stop();
+
+    REQUIRE(a_overflows.load() > 0);
+    REQUIRE(b_overflows.load() == 0);
+}
+
+TEST_CASE("interrupt node overflow callback fires on full output", "[interrupt_node][overflow]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    g_interrupt_counter.store(0);
+    auto node = make_interrupt_node<interrupt_produce>(pool, out<>{});
+
+    Channel<int> full_ch(1);
+    full_ch.push(99);
+    node.set_output_channel<0>(&full_ch);
+
+    std::atomic<int> overflow_count{0};
+    node.set_overflow_callback([&](auto) { overflow_count.fetch_add(1); });
+
+    node.start();
+    auto trigger = node.get_trigger();
+    trigger(); trigger(); trigger();
+
+    std::this_thread::sleep_for(std::chrono::milliseconds(50));
+    node.stop();
+    pool->stop();
+
+    REQUIRE(overflow_count.load() > 0);
+}
+
+// ── self_stop: disable inputs + outputs on crash ──────────────────────────────
+
+static int always_throw(int) { throw std::runtime_error("node crashed"); return 0; }
+
+TEST_CASE("pool node self_stop disables output on crash so downstream sees closed", "[pool_node][self_stop]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    auto node = make_pool_node<always_throw>(pool, 5);
+    Channel<int> out_ch(10);
+    node.set_output_channel<0>(&out_ch);
+
+    node.start();
+    node.input_channel<0>().push(1);
+
+    std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+    REQUIRE_FALSE(out_ch.is_accepting());
+
+    node.stop();
+    pool->stop();
+}
+
+TEST_CASE("pool node self_stop disables input on crash", "[pool_node][self_stop]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    auto node = make_pool_node<always_throw>(pool, 5);
+    Channel<int> out_ch(5);
+    node.set_output_channel<0>(&out_ch);
+
+    node.start();
+    node.input_channel<0>().push(1);
+
+    std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+    REQUIRE_FALSE(node.input_channel<0>().is_accepting());
+
+    node.stop();
+    pool->stop();
+}
+
+TEST_CASE("pool node closed callback fires on self_stop from crash", "[pool_node][self_stop]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    auto node = make_pool_node<always_throw>(pool, 5);
+    Channel<int> out_ch(5);
+    node.set_output_channel<0>(&out_ch);
+
+    std::atomic<bool> closed_fired{false};
+    node.set_closed_callback([&](auto) { closed_fired.store(true); });
+
+    node.start();
+    node.input_channel<0>().push(1);
+
+    std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+    REQUIRE(closed_fired.load());
+
+    node.stop();
+    pool->stop();
+}
+
+// ── Network-level event callbacks ─────────────────────────────────────────────
+
+TEST_CASE("network_overflow_callback fires on overflow", "[pool_node][network]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    auto node = make_pool_node<double_it>(pool);
+
+    Channel<int> full_ch(1);
+    full_ch.push(0);
+    node.set_output_channel<0>(&full_ch);
+
+    std::atomic<int> net_overflows{0};
+    node.set_network_overflow_callback([&](auto) { net_overflows.fetch_add(1); });
+
+    node.start();
+    node.input_channel<0>().push(1);
+    node.input_channel<0>().push(2);
+
+    std::this_thread::sleep_for(std::chrono::milliseconds(50));
+    node.stop();
+    pool->stop();
+
+    REQUIRE(net_overflows.load() > 0);
+}
+
+TEST_CASE("network_closed_callback fires on crash", "[pool_node][network]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    auto node = make_pool_node<always_throw>(pool);
+    Channel<int> out_ch(5);
+    node.set_output_channel<0>(&out_ch);
+
+    std::atomic<bool> net_closed{false};
+    node.set_network_closed_callback([&](auto) { net_closed.store(true); });
+
+    node.start();
+    node.input_channel<0>().push(1);
+
+    std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+    REQUIRE(net_closed.load());
+
+    node.stop();
+    pool->stop();
+}
+
+TEST_CASE("per-node and network overflow callbacks both fire independently", "[pool_node][network]") {
+    auto pool = std::make_shared<ThreadPool>(2);
+    pool->start();
+
+    auto node = make_pool_node<double_it>(pool);
+
+    Channel<int> full_ch(1);
+    full_ch.push(0);
+    node.set_output_channel<0>(&full_ch);
+
+    std::atomic<int> per_node{0}, network{0};
+    node.set_overflow_callback([&](auto) { per_node.fetch_add(1); });
+    node.set_network_overflow_callback([&](auto) { network.fetch_add(1); });
+
+    node.start();
+    node.input_channel<0>().push(1);
+    node.input_channel<0>().push(2);
+
+    std::this_thread::sleep_for(std::chrono::milliseconds(50));
+    node.stop();
+    pool->stop();
+
+    REQUIRE(per_node.load() > 0);
+    REQUIRE(network.load() > 0);
+}

tests/test_scheduler.cpp

@@ -0,0 +1,229 @@
+#include <catch2/catch_test_macros.hpp>
+#include <kpn/scheduler.hpp>
+
+#include <atomic>
+#include <chrono>
+#include <thread>
+#include <vector>
+#include <mutex>
+
+using namespace kpn;
+using namespace std::chrono_literals;
+
+// ── basic execution ───────────────────────────────────────────────────────────
+
+TEST_CASE("scheduler runs submitted tasks", "[scheduler]") {
+    ThreadPool pool(2);
+    pool.start();
+
+    std::atomic<int> counter{0};
+    for (int i = 0; i < 100; ++i)
+        pool.submit([&counter]{ counter.fetch_add(1, std::memory_order_relaxed); });
+
+    pool.drain();
+    REQUIRE(counter.load() == 100);
+    pool.stop();
+}
+
+TEST_CASE("scheduler single thread executes all tasks", "[scheduler]") {
+    ThreadPool pool(1);
+    pool.start();
+
+    std::atomic<int> counter{0};
+    for (int i = 0; i < 50; ++i)
+        pool.submit([&counter]{ counter.fetch_add(1, std::memory_order_relaxed); });
+
+    pool.drain();
+    REQUIRE(counter.load() == 50);
+    pool.stop();
+}
+
+// ── drain ─────────────────────────────────────────────────────────────────────
+
+TEST_CASE("drain returns immediately when pool is idle", "[scheduler]") {
+    ThreadPool pool(2);
+    pool.start();
+    pool.drain();   // nothing submitted — should return immediately
+    pool.stop();
+}
+
+TEST_CASE("drain waits for all tasks to complete", "[scheduler]") {
+    ThreadPool pool(4);
+    pool.start();
+
+    std::atomic<int> counter{0};
+    constexpr int N = 200;
+
+    for (int i = 0; i < N; ++i) {
+        pool.submit([&counter]{
+            std::this_thread::sleep_for(1ms);
+            counter.fetch_add(1, std::memory_order_relaxed);
+        });
+    }
+
+    pool.drain();
+    REQUIRE(counter.load() == N);
+    pool.stop();
+}
+
+TEST_CASE("drain is safe to call multiple times", "[scheduler]") {
+    ThreadPool pool(2);
+    pool.start();
+
+    std::atomic<int> counter{0};
+    pool.submit([&counter]{ counter.fetch_add(1, std::memory_order_relaxed); });
+    pool.drain();
+    REQUIRE(counter.load() == 1);
+
+    pool.submit([&counter]{ counter.fetch_add(1, std::memory_order_relaxed); });
+    pool.drain();
+    REQUIRE(counter.load() == 2);
+
+    pool.stop();
+}
+
+// ── priority ordering ─────────────────────────────────────────────────────────
+
+TEST_CASE("higher priority tasks run before lower priority on single thread", "[scheduler]") {
+    // Single thread guarantees serial execution — we can observe order.
+    ThreadPool pool(1);
+    pool.start();
+
+    // Pause the worker so we can fill the queue before it drains.
+    std::mutex gate;
+    gate.lock();
+    pool.submit([&gate]{ std::lock_guard lg(gate); }); // blocks worker
+
+    std::vector<float> order;
+    std::mutex order_mx;
+
+    for (float p : {0.1f, 0.9f, 0.5f, 0.8f, 0.2f}) {
+        pool.submit([p, &order, &order_mx]{
+            std::lock_guard lg(order_mx);
+            order.push_back(p);
+        }, p);
+    }
+
+    gate.unlock();   // release the blocking task
+    pool.drain();
+    pool.stop();
+
+    // order should be descending by priority
+    REQUIRE(order.size() == 5);
+    for (std::size_t i = 1; i < order.size(); ++i)
+        REQUIRE(order[i - 1] >= order[i]);
+}
+
+TEST_CASE("equal priority tasks execute in FIFO order on single thread", "[scheduler]") {
+    ThreadPool pool(1);
+    pool.start();
+
+    std::mutex gate;
+    gate.lock();
+    pool.submit([&gate]{ std::lock_guard lg(gate); });
+
+    std::vector<int> order;
+    std::mutex order_mx;
+
+    for (int i = 0; i < 5; ++i) {
+        pool.submit([i, &order, &order_mx]{
+            std::lock_guard lg(order_mx);
+            order.push_back(i);
+        }, 0.5f);   // all same priority
+    }
+
+    gate.unlock();
+    pool.drain();
+    pool.stop();
+
+    REQUIRE(order == std::vector<int>{0, 1, 2, 3, 4});
+}
+
+// ── total_ / active_ accounting ───────────────────────────────────────────────
+
+TEST_CASE("snapshot queue depth and active counts are consistent", "[scheduler]") {
+    ThreadPool pool(2);
+    pool.start();
+
+    // While tasks are running, active should be > 0 and total >= active.
+    std::atomic<bool> running{false};
+    std::mutex gate;
+    gate.lock();
+
+    for (int i = 0; i < 4; ++i) {
+        pool.submit([&gate, &running]{
+            running.store(true, std::memory_order_relaxed);
+            std::lock_guard lg(gate);
+        });
+    }
+
+    // Spin until at least one task has started
+    while (!running.load(std::memory_order_relaxed))
+        std::this_thread::yield();
+
+    auto snap = pool.snapshot("test");
+    REQUIRE(snap.active_count > 0);
+    REQUIRE(snap.queue_depth + snap.active_count > 0);
+
+    gate.unlock();
+    pool.drain();
+
+    auto snap2 = pool.snapshot("test");
+    REQUIRE(snap2.active_count == 0);
+    REQUIRE(snap2.queue_depth  == 0);
+
+    pool.stop();
+}
+
+TEST_CASE("submitted and completed counters are accurate", "[scheduler]") {
+    ThreadPool pool(3);
+    pool.start();
+
+    constexpr int N = 60;
+    for (int i = 0; i < N; ++i)
+        pool.submit([]{ std::this_thread::yield(); });
+
+    pool.drain();
+    auto snap = pool.snapshot("test");
+    REQUIRE(snap.tasks_submitted == static_cast<uint64_t>(N));
+    REQUIRE(snap.tasks_completed == static_cast<uint64_t>(N));
+
+    pool.stop();
+}
+
+// ── work stealing ─────────────────────────────────────────────────────────────
+
+TEST_CASE("work stealing: all tasks complete with uneven initial distribution", "[scheduler]") {
+    // 4-thread pool. Submit a burst to ensure some threads start empty and must steal.
+    ThreadPool pool(4);
+    pool.start();
+
+    std::atomic<int> counter{0};
+    constexpr int N = 400;
+
+    for (int i = 0; i < N; ++i)
+        pool.submit([&counter]{
+            std::this_thread::sleep_for(100us);
+            counter.fetch_add(1, std::memory_order_relaxed);
+        });
+
+    pool.drain();
+    REQUIRE(counter.load() == N);
+    pool.stop();
+}
+
+TEST_CASE("work stealing: tasks complete with more threads than initial queue targets", "[scheduler]") {
+    // With round-robin, some threads may get no tasks initially and must steal.
+    constexpr std::size_t THREADS = 8;
+    ThreadPool pool(THREADS);
+    pool.start();
+
+    std::atomic<int> counter{0};
+    // Submit fewer tasks than threads so most threads must steal
+    for (int i = 0; i < 4; ++i)
+        pool.submit([&counter]{ counter.fetch_add(1, std::memory_order_relaxed); });
+
+    pool.drain();
+    REQUIRE(counter.load() == 4);
+    pool.stop();
+}

tests/test_shared_resource.cpp

@@ -0,0 +1,239 @@
+#include <catch2/catch_test_macros.hpp>
+#include <catch2/catch_approx.hpp>
+#include <kpn/shared_resource.hpp>
+#include <kpn/channel.hpp>
+
+#include <atomic>
+#include <chrono>
+#include <thread>
+#include <vector>
+
+using namespace kpn;
+using namespace std::chrono_literals;
+
+// ── Basic acquire / release ───────────────────────────────────────────────────
+
+TEST_CASE("acquire returns guard that accesses the resource", "[shared_resource]") {
+    SharedResource<int> res(42);
+    {
+        auto g = res.acquire();
+        REQUIRE(*g == 42);
+        *g = 99;
+    } // g released here — second acquire must not overlap in the same thread
+    {
+        auto g2 = res.acquire();
+        REQUIRE(*g2 == 99);
+    }
+}
+
+TEST_CASE("guard operator-> reaches resource members", "[shared_resource]") {
+    struct Pair { int x{1}; int y{2}; };
+    SharedResource<Pair> res;
+    auto g = res.acquire();
+    REQUIRE(g->x == 1);
+    REQUIRE(g->y == 2);
+    g->x = 10;
+    REQUIRE(g.get().x == 10);
+}
+
+TEST_CASE("guard releases on scope exit", "[shared_resource]") {
+    SharedResource<int> res(0);
+    {
+        auto g = res.acquire();
+        REQUIRE(res.snapshot("r").held);
+    }
+    // After guard destroyed, resource is free
+    REQUIRE(!res.snapshot("r").held);
+}
+
+// ── Mutual exclusion ──────────────────────────────────────────────────────────
+
+TEST_CASE("only one thread holds the resource at a time", "[shared_resource]") {
+    SharedResource<int> res(0);
+    std::atomic<int> concurrent_holders{0};
+    std::atomic<int> violations{0};
+    std::atomic<bool> go{false};
+
+    auto worker = [&] {
+        while (!go.load()) std::this_thread::yield();
+        for (int i = 0; i < 20; ++i) {
+            auto g = res.acquire();
+            int h = concurrent_holders.fetch_add(1) + 1;
+            if (h > 1) violations.fetch_add(1);
+            std::this_thread::sleep_for(100us);
+            concurrent_holders.fetch_sub(1);
+        }
+    };
+
+    std::vector<std::thread> threads;
+    for (int i = 0; i < 4; ++i) threads.emplace_back(worker);
+    go.store(true);
+    for (auto& t : threads) t.join();
+
+    REQUIRE(violations.load() == 0);
+}
+
+// ── Priority ordering ─────────────────────────────────────────────────────────
+
+TEST_CASE("higher priority waiter is served before lower priority waiter", "[shared_resource]") {
+    SharedResource<int> res(0);
+
+    // Hold the resource so threads have to queue.
+    auto holder = res.acquire();
+
+    std::vector<int> order;
+    std::mutex order_mtx;
+
+    // Launch two waiters: low priority first, then high priority.
+    std::thread low([&] {
+        auto g = res.acquire([] { return 0.1f; });
+        std::lock_guard lk(order_mtx);
+        order.push_back(1);
+    });
+    std::this_thread::sleep_for(5ms); // ensure low is queued first
+
+    std::thread high([&] {
+        auto g = res.acquire([] { return 0.9f; });
+        std::lock_guard lk(order_mtx);
+        order.push_back(2);
+    });
+    std::this_thread::sleep_for(5ms); // ensure high is also queued
+
+    // Release — high priority should win even though low arrived first.
+    { auto drop = std::move(holder); }
+
+    low.join();
+    high.join();
+
+    REQUIRE(order.size() == 2);
+    REQUIRE(order[0] == 2); // high priority served first
+    REQUIRE(order[1] == 1);
+}
+
+// ── acquire_balanced uses channel fills ───────────────────────────────────────
+
+TEST_CASE("acquire_balanced: full input + empty output gives score ~1.0", "[shared_resource]") {
+    // We test the priority function indirectly via ordering.
+    // Node A: in=full, out=empty  → score ≈ 1.0  (high)
+    // Node B: in=empty, out=full  → score ≈ 0.0  (low)
+
+    Channel<int> in_a(4);   // fill it
+    Channel<int> out_a(4);  // leave empty
+    Channel<int> in_b(4);   // leave empty
+    Channel<int> out_b(4);  // fill it
+
+    in_a.enable();  out_a.enable();
+    in_b.enable();  out_b.enable();
+
+    for (int i = 0; i < 4; ++i) { in_a.push(i); out_b.push(i); }
+
+    SharedResource<int> res(0);
+    auto holder = res.acquire(); // block others
+
+    std::vector<int> order;
+    std::mutex mtx;
+
+    // Node B (low priority) waits first
+    std::thread tb([&] {
+        auto g = res.acquire_balanced(in_b, out_b);
+        std::lock_guard lk(mtx);
+        order.push_back(2);
+    });
+    std::this_thread::sleep_for(5ms);
+
+    // Node A (high priority) waits second
+    std::thread ta([&] {
+        auto g = res.acquire_balanced(in_a, out_a);
+        std::lock_guard lk(mtx);
+        order.push_back(1);
+    });
+    std::this_thread::sleep_for(5ms);
+
+    { auto drop = std::move(holder); } // release
+
+    ta.join();
+    tb.join();
+
+    REQUIRE(order.size() == 2);
+    REQUIRE(order[0] == 1); // node A served first despite arriving second
+}
+
+// ── Statistics ────────────────────────────────────────────────────────────────
+
+TEST_CASE("stats: acquisitions counted correctly", "[shared_resource]") {
+    SharedResource<int> res(0);
+    {
+        auto g1 = res.acquire();
+    }
+    {
+        auto g2 = res.acquire();
+    }
+    REQUIRE(res.snapshot("r").acquisitions == 2);
+}
+
+TEST_CASE("stats: peak_waiters reflects maximum concurrent queue depth", "[shared_resource]") {
+    SharedResource<int> res(0);
+    auto holder = res.acquire();
+
+    std::atomic<int> ready{0};
+
+    auto waiter = [&] {
+        ready.fetch_add(1);
+        auto g = res.acquire();
+    };
+
+    std::thread t1(waiter), t2(waiter), t3(waiter);
+
+    // Wait until all three are queued
+    while (ready.load() < 3) std::this_thread::sleep_for(1ms);
+    std::this_thread::sleep_for(5ms); // give them time to block on acquire
+
+    { auto drop = std::move(holder); } // release
+
+    t1.join(); t2.join(); t3.join();
+
+    REQUIRE(res.snapshot("r").peak_waiters >= 2); // at least 2 queued simultaneously
+}
+
+TEST_CASE("stats: current_waiters returns to 0 after all served", "[shared_resource]") {
+    SharedResource<int> res(0);
+    auto holder = res.acquire();
+
+    std::thread t1([&] { auto g = res.acquire(); });
+    std::thread t2([&] { auto g = res.acquire(); });
+    std::this_thread::sleep_for(10ms);
+
+    { auto drop = std::move(holder); }
+    t1.join(); t2.join();
+
+    REQUIRE(res.snapshot("r").current_waiters == 0);
+}
+
+TEST_CASE("stats: avg_wait_ms is positive when contention occurred", "[shared_resource]") {
+    SharedResource<int> res(0);
+    {
+        auto holder = res.acquire();
+        std::thread t([&] { auto g = res.acquire(); });
+        std::this_thread::sleep_for(10ms);
+        { auto drop = std::move(holder); }
+        t.join();
+    }
+    REQUIRE(res.snapshot("r").avg_wait_ms > 0.0);
+}
+
+// ── No-arg acquire ────────────────────────────────────────────────────────────
+
+TEST_CASE("no-arg acquire works and releases correctly", "[shared_resource]") {
+    SharedResource<int> res(7);
+    auto g = res.acquire();
+    REQUIRE(*g == 7);
+    REQUIRE(res.snapshot("r").held);
+}
+
+// ── make_shared_resource factory ──────────────────────────────────────────────
+
+TEST_CASE("make_shared_resource constructs with forwarded args", "[shared_resource]") {
+    auto res = make_shared_resource<std::string>("hello");
+    auto g = res.acquire();
+    REQUIRE(*g == "hello");
+}