Fix http handler not found: pre-register via el_runtime_register_handler

elb links without -rdynamic so dlsym(RTLD_DEFAULT, "handle_request")
returns NULL at runtime. http_set_handler stores the name as active but
never finds a function pointer, causing every request to return
"el-runtime: no http handler registered" even after http_serve is called.

Fix: add a __attribute__((constructor)) in web_stubs.c that calls
el_runtime_register_handler("handle_request", handle_request) directly,
bypassing dlsym entirely. The handler is in the registry before main()
runs, so http_lookup_active() finds it on the first request.

.gitea/workflows/stage.yaml

@@ -190,18 +190,34 @@ jobs:
              -lcurl -lpthread -ldl -lm -lssl -lcrypto
           echo "soul-demo compiled: $(ls -lh dist/soul-demo)"
           # Package as minimal OCI image for k3s import
-          docker build -f dist/Dockerfile.soul-demo -t soul-demo:local dist/
+          # --no-cache: prevents reuse of corrupted overlay2 layers from prior failed runs
+          docker build --no-cache -f dist/Dockerfile.soul-demo -t soul-demo:local dist/
           docker save soul-demo:local -o dist/soul-demo-image.tar
           echo "soul-demo-image.tar: $(du -sh dist/soul-demo-image.tar | cut -f1)"
           docker rmi soul-demo:local 2>/dev/null || true
 
+      - name: Download k3s binary
+        # Pre-download k3s on the host runner so Dockerfile.stage can COPY it
+        # directly. Previously k3s was downloaded inside the Docker builder stage,
+        # which combined with build-essential and C compilation caused RWLayer nil
+        # corruption on the runner's overlay2 driver. Host-runner download is safe.
+        if: steps.changetype.outputs.asset_only != 'true'
+        run: |
+          set -euo pipefail
+          curl -fL --retry 3 --retry-delay 10 \
+            https://github.com/k3s-io/k3s/releases/download/v1.32.4%2Bk3s1/k3s \
+            -o dist/k3s
+          chmod +x dist/k3s
+          echo "k3s: $(ls -lh dist/k3s)"
+
       - name: Build and tag image
         if: steps.changetype.outputs.asset_only != 'true'
         run: |
           set -euo pipefail
+          # --no-cache: prevents reuse of corrupted overlay2 layers from prior failed runs.
+          # Dockerfile.stage is now single-stage (no builder) so build is fast even without cache.
           docker build \
-            --build-arg BUILDKIT_INLINE_CACHE=1 \
+            --no-cache \
-            --cache-from us-central1-docker.pkg.dev/neuron-785695/neuron-marketing/marketing:stage-latest \
             -f Dockerfile.stage \
             -t "marketing:${{ steps.tag.outputs.tag }}" \
             .

Dockerfile.stage

@@ -4,56 +4,24 @@
 #   - neuron-web  on port 8080 (landing page server)
 #   - soul-demo   on port 7772 (demo chat, localhost only)
 #
-# neuron-web is built by `elb build` in CI (not here). elb compiles each
+# All binaries (neuron-web, soul-demo, k3s) are pre-built by CI on the host
-# .el source independently and links the result — no combined mega-file,
+# runner before this Dockerfile runs. This keeps the Docker build single-stage
-# no exponential memory growth. The binary lands at dist/neuron-landing
+# with no compilation and no network downloads, eliminating the multi-stage
-# (linux/amd64) and is COPY'd directly into the runtime image.
+# complexity that caused RWLayer corruption on the runner's overlay2 driver.
 #
-# soul-demo.c is pre-committed (small, no OOM risk) and compiled here.
+# CI pre-build steps (in stage.yaml):
+#   - neuron-web: built by `elb build` → dist/neuron-landing
+#   - soul-demo:  compiled by cc on host → dist/soul-demo
+#   - k3s:        downloaded by curl on host → dist/k3s
 
-# ── Stage 1: compile soul-demo ────────────────────────────────────────────────
+FROM ubuntu:24.04
-FROM debian:bookworm-slim AS builder
+
+ENV DEBIAN_FRONTEND=noninteractive
 
 RUN apt-get update \
  && apt-get install -y --no-install-recommends \
-        build-essential \
+        libcurl4t64 \
-        curl \
+        libssl3t64 \
-        libcurl4-openssl-dev \
-        libssl-dev \
-        ca-certificates \
- && rm -rf /var/lib/apt/lists/*
-
-WORKDIR /build
-
-COPY runtime/el_runtime.c runtime/el_runtime.h ./
-
-# Pre-compile el_runtime as a separate cached layer.
-# el_runtime.c changes rarely; main.c changes every run.
-# Splitting this out means el_runtime.o is cached across builds when only main.c changes.
-# -DHAVE_CURL: the staged el_runtime.c (from el.git) guards the OTLP observability
-# section (emit_metric, emit_log, trace_span_*) behind #ifdef HAVE_CURL.
-# libcurl IS installed above, so define HAVE_CURL to enable those functions.
-RUN cc -O2 -DHAVE_CURL -c el_runtime.c -I. -o el_runtime.o
-
-COPY dist/soul-demo.c dist/vessel_stubs.c ./
-
-RUN cc -O2 -rdynamic \
-       -DEL_SOUL_DEMO_BUILD \
-       -o soul-demo \
-       soul-demo.c vessel_stubs.c el_runtime.o \
-       -lcurl -lpthread -ldl -lm -lssl -lcrypto
-
-# ── Download k3s binary ───────────────────────────────────────────────────────
-RUN curl -fL --retry 3 --retry-delay 10 https://github.com/k3s-io/k3s/releases/download/v1.32.4%2Bk3s1/k3s -o /usr/local/bin/k3s \
- && chmod +x /usr/local/bin/k3s
-
-# ── Stage 2: runtime image ────────────────────────────────────────────────────
-FROM debian:bookworm-slim
-
-RUN apt-get update \
- && apt-get install -y --no-install-recommends \
-        libcurl4 \
-        libssl3 \
         ca-certificates \
  && rm -rf /var/lib/apt/lists/* \
  && groupadd -r landing && useradd -r -g landing landing \
@@ -67,10 +35,13 @@ RUN apt-get update \
 COPY dist/neuron-landing /usr/local/bin/neuron-web
 RUN chmod +x /usr/local/bin/neuron-web
 
-COPY --from=builder /build/soul-demo /usr/local/bin/soul-demo
+# soul-demo binary — compiled by cc on host runner in CI
+COPY dist/soul-demo /usr/local/bin/soul-demo
+RUN chmod +x /usr/local/bin/soul-demo
 
-# k3s binary
+# k3s binary — downloaded from GitHub releases by CI
-COPY --from=builder /usr/local/bin/k3s /usr/local/bin/k3s
+COPY dist/k3s /usr/local/bin/k3s
+RUN chmod +x /usr/local/bin/k3s
 
 # soul-demo OCI image tar — k3s imports this at startup (no registry needed)
 RUN mkdir -p /var/lib/rancher/k3s/agent/images

dist/entrypoint.sh

@@ -16,28 +16,26 @@ echo "[entrypoint] Starting k3s server (embedded soul-demo orchestrator)..."
 # --disable metrics-server: saves ~50MB RAM
 # --write-kubeconfig-mode=644: allow non-root reads
 # --data-dir: use the pre-chowned dir
+# --flannel-iface=eth0: explicitly set the network interface.
+#   Cloud Run gen2 provides eth0 but k3s default IP detection walks the routing
+#   table looking for a default route, which fails in Cloud Run's network sandbox.
+#   Pinning to eth0 bypasses that detection and lets k3s bind correctly.
 k3s server \
   --disable traefik \
   --disable servicelb \
   --disable metrics-server \
   --write-kubeconfig-mode=644 \
   --data-dir /var/lib/rancher/k3s \
-  --node-name soul-node &
+  --node-name soul-node \
+  --flannel-iface=eth0 &
 
 K3S_PID=$!
 
-echo "[entrypoint] Waiting for k3s to become ready..."
+# Start neuron-web immediately — do NOT block on k3s becoming ready.
-until k3s kubectl get nodes --no-headers 2>/dev/null | grep -q "Ready"; do
+# Cloud Run's startup probe requires port 8080 to be listening within the
-  sleep 2
+# startup timeout. k3s may take 30-60s to initialise; blocking here causes
-done
+# probe failures and container termination before neuron-web ever starts.
-echo "[entrypoint] k3s ready. soul-demo Deployment will be applied automatically from manifests."
+# soul-demo becomes available asynchronously once k3s is ready. neuron-web
-
+# handles soul-demo being temporarily unavailable gracefully.
-# Wait for soul-demo pod to be Running before starting neuron-web
+echo "[entrypoint] Starting neuron-web on port ${PORT:-8080} (k3s initialising in background)..."
-echo "[entrypoint] Waiting for soul-demo pod..."
-until k3s kubectl get pods -l app=soul-demo --no-headers 2>/dev/null | grep -q "Running"; do
-  sleep 3
-done
-echo "[entrypoint] soul-demo is running."
-
-echo "[entrypoint] Starting neuron-web on port ${PORT:-8080}..."
 exec /usr/local/bin/neuron-web

dist/web_stubs.c

@@ -6,6 +6,31 @@
 #include <unistd.h>
 #include "el_runtime.h"
 
+/* Pre-register the El HTTP handler so it is found by http_lookup_active()
+ * regardless of whether the binary was linked with -rdynamic.
+ *
+ * el_runtime's http_set_handler resolves handler names via:
+ *   dlsym(RTLD_DEFAULT, "handle_request")
+ * but dlsym only searches the dynamic symbol table, which only contains
+ * user-defined symbols when the executable is linked with -rdynamic.
+ * elb does not add -rdynamic, so dlsym returns NULL and routes return
+ * "el-runtime: no http handler registered" even though http_serve is called.
+ *
+ * The fix: forward-declare handle_request here and register it directly
+ * via el_runtime_register_handler before main() runs. This populates the
+ * handler registry so http_lookup_active() finds it without needing dlsym.
+ */
+extern el_val_t handle_request(el_val_t method, el_val_t path, el_val_t body);
+/* el_runtime_register_handler is intentionally not declared in el_runtime.h
+ * ("extern lookup works since C symbols are global" — runtime comment). */
+extern void el_runtime_register_handler(const char* name,
+                                        el_val_t (*fn)(el_val_t, el_val_t, el_val_t));
+
+__attribute__((constructor))
+static void pre_register_http_handlers(void) {
+    el_runtime_register_handler("handle_request", handle_request);
+}
+
 el_val_t http_get_auth(el_val_t url, el_val_t tok) {
     char bearer[2048]; snprintf(bearer, sizeof(bearer), "Bearer %s", EL_CSTR(tok));
     el_val_t hdr_val = EL_STR(bearer);