…lobal args

extract_plugin_config() read from NodeOptions::parameter_overrides() which
is always empty for --params-file YAML params (they go to the rcl global
context instead). Plugins received empty config and fell back to defaults.

Fix: declare_plugin_params_from_yaml() accesses rcl_arguments_get_param_overrides()
from the global context, finds params matching the plugin prefix, and declares
them on the node with their typed values. extract_plugin_config() then reads
them via list_parameters()/get_parameter().

The NodeOptions::parameter_overrides path is preserved as the primary source
for programmatic overrides (used in unit tests).

… test

The previous test created a full GatewayNode which hung on Humble/Rolling CI
due to heavy DDS/HTTP initialization. Replace with a plain rclcpp::Node that
directly tests the declare_plugin_params_from_yaml() logic:

1. Writes YAML with plugin params (string, int, bool, double)
2. Inits rclcpp with --params-file
3. Confirms parameter_overrides() is empty (proves the bug)
4. Calls declare_plugin_params_from_yaml() and verifies all types resolve

Test runs in ~50ms instead of timing out.

…header

- Move function to param_utils.hpp so test uses the real implementation
  instead of a diverging copy
- Use unique_ptr RAII guard for rcl_params_t instead of manual cleanup
- Log param name on declaration failure instead of silent catch(...)
- Warn when encountering unsupported array param types

Commit 5224420 removed libcpp-httplib-dev from rosdep skip-keys thinking
the vendored cpp-httplib fallback would cover it. That's true at the CMake
level, but rosdep runs first and fails on humble (Ubuntu 22.04) because
libcpp-httplib-dev is not available in the apt repo.

Add it back to skip-keys so rosdep skips the lookup on every distro; the
CMake macro medkit_find_cpp_httplib then uses the vendored header in
src/ros2_medkit_gateway/src/vendored/cpp_httplib/ everywhere. Single
fallback path for humble/jazzy/rolling = consistent builds.

Fixes the build-and-push (humble) job failure on main.

Two race conditions in sample_topic() called from httplib threads while
the executor spins on the main thread:

1. create_generic_subscription from httplib thread races with the
   executor iterating callback groups. Fix: dedicated
   MutuallyExclusiveCallbackGroup for sampling subscriptions isolates
   them from the default group. A sampling_mutex_ serializes subscription
   lifecycle so only one sampling subscription is active at a time.

2. Callback lambda captures &message_promise and &received by reference
   to stack-local variables. On timeout, the function returns and destroys
   the stack while the executor may still be dispatching the callback.
   Fix: heap-allocate shared state (SampleState) behind a shared_ptr so
   the callback keeps state alive regardless of the caller's lifetime.

Adds 3 thread-safety tests:
- SampleTopicTimeoutDoesNotCrash (20 rapid timeouts)
- ConcurrentSampleTopicDoesNotCrash (4 threads x 5 iterations)
- RapidTimeoutWithActivePublisher (30 boundary-condition iterations)

Manifested as non-deterministic gateway SIGSEGV (exit -11) during
test_data_read on rolling CI.

…ublisher

When a data trigger was created before its topic was discoverable (late
publisher scenario), resolved_topic_name was empty and TriggerManager
skipped calling topic_subscriber_->subscribe(). The trigger appeared
"active" but had no backing subscription - no SSE events were ever
delivered.

Add deferred re-resolution:
- TriggerManager stores unresolved data triggers at creation time
- GatewayNode wires a ResolveTopicFn (cache lookup by entity_id +
  resource_path suffix matching)
- TriggerTopicSubscriber's 5s retry timer calls
  TriggerManager::retry_unresolved_triggers()
- Resolved triggers get their topic_name updated and subscribed

Adds 3 unit tests for deferred resolution safety.

Fixes test_triggers_late_publisher flaking on humble where DDS
discovery is slower than on jazzy/rolling.

Commit f00a156 added callback_mutex_ synchronization to rpm_sensor as
an experiment but never followed up with the other demo nodes. On humble,
TimerBase::cancel() does not wait for in-flight callbacks; if the
destructor resets the publisher while a callback is running, the callback
dereferences a null SharedPtr causing SIGSEGV (exit -11).

Apply the same pattern to all 7 remaining demo nodes with timer+publisher:
- beacon_publisher, brake_actuator, brake_pressure_sensor,
  door_status_sensor, engine_temp_sensor, lidar_sensor, light_controller

Pattern: callback_mutex_ locked in every timer callback (with null-check
on publisher) and in the destructor after cancel() before resetting members.

lidar_sensor required extracting check_and_report_faults_unlocked() to
avoid deadlock when on_parameter_change (holding lock) calls fault check.

Fixes recurring humble SIGSEGV in test_entity_routing, test_data_read,
test_entity_model_runtime, and any test using demo nodes.

… API stubs

Two build issues introduced on main by the VDA 5050 merge:

1. medkit_find_cpp_httplib() called without VENDORED_DIR - fails on
   humble where libcpp-httplib-dev is unavailable. Point to gateway's
   vendored copy like graph_provider does.

2. Test stubs used old PluginContext API (httplib types). Update to
   PluginRequest/PluginResponse wrapper types.

…ption creation

SingleThreadedExecutor does not synchronize Node::create_generic_subscription()
called from external threads (cpp-httplib handlers) with its internal wait set
iteration. This causes non-deterministic SIGSEGV on rolling where rclcpp
internals are less forgiving of concurrent node manipulation.

MultiThreadedExecutor properly synchronizes subscription lifecycle across
threads. All gateway callbacks are already protected by mutexes (EntityCache,
LogManager, TriggerManager, TriggerTopicSubscriber), so concurrent callback
execution is safe.

This is the definitive fix for the rolling SIGSEGV in test_data_read.
The previous sampling_mutex_ + MutuallyExclusiveCallbackGroup fix reduced
the race window but did not eliminate it.