Merge pipeline architecture improvements

[bburda]

Summary

Several architecture improvements for the merge pipeline to reduce lock contention, improve extensibility, and fix subtle merge semantics.

---

Proposed solution (optional)

1. Double-buffering for HybridDiscovery mutex

HybridDiscoveryStrategy::refresh() holds the mutex during the entire pipeline_.execute() call, which includes synchronous ROS 2 graph introspection. All HTTP handlers calling discover_*() block during refresh, causing request latency spikes.

File: src/ros2_medkit_gateway/src/discovery/hybrid_discovery.cpp:48-55

Fix: Execute pipeline without lock, then swap result under lock.

2. Replace dynamic_cast for RuntimeLayer statistics

MergePipeline::execute() uses dynamic_cast<RuntimeLayer*> to get last_filtered_count(). This couples the pipeline to a concrete type, violating OCP.

File: src/ros2_medkit_gateway/src/discovery/merge_pipeline.cpp:418

Fix: Add virtual size_t filtered_count() const { return 0; } to DiscoveryLayer base class.

3. Fix merge_bool sticky OR for external field

merge_bool() uses target = target || source for MergeWinner::BOTH. Once any layer sets a boolean to true, it can never be reverted. Correct for is_online but wrong for external.

File: src/ros2_medkit_gateway/src/discovery/merge_pipeline.cpp:89-103

Fix: Move external to METADATA field group where manifest is AUTHORITATIVE by default.

4. Make MergePipeline accessors private

get_last_report() and get_linking_result() are public but only called through HybridDiscoveryStrategy. Direct calls would be a data race (no internal mutex).

File: src/ros2_medkit_gateway/include/ros2_medkit_gateway/discovery/merge_pipeline.hpp:67-85

Fix: Make them private with friend class HybridDiscoveryStrategy.

---

Additional context (optional)

Found during self-review of PR #258. Item 1 (mutex contention) is the most impactful for production - it affects request latency during every discovery refresh cycle.