fix: break circular GObject references that leaked pipelines and sockets

[srperens]

Summary

• Strong references to pipeline and elements captured in connect_pad_added closures in linking.rs created a circular reference chain: pipeline owns elements → elements own signal handler closures → closures own pipeline+elements. This prevented GStreamer from ever finalizing the pipeline, leaking all OS resources (UDP sockets, threads, file descriptors) on every pipeline restart.
• On production, this accumulated ~5870 leaked UDP sockets after repeated flow restarts, causing kernel socket buffer overflow and AES67 packet loss (21% loss).
• Replaced strong refs with WeakRef in dynamic pad handler closures, balanced add_signal_watch/remove_signal_watch ref-counting in bus cleanup, and added permanent runtime verification + regression test.

Changes

• linking.rs — Replace pipeline.clone() and elements.clone() with WeakRef in setup_dynamic_pad_handlers() closures
• bus.rs — Balance remove_signal_watch calls to match add_signal_watch count
• mod.rs — Clean up Drop impl ordering (stop QoS task before set_state)
• state.rs — Add runtime leak detection in stop_flow(): logs ERROR if pipeline/elements survive after PipelineManager drop
• state.rs — Add flow_name(), element_weak_refs() accessors
• CLAUDE.md — Add rule: never capture strong GStreamer refs in signal handler closures
• pipeline_lifecycle_test.rs — Permanent regression test using real PipelineManager

Test plan

• cargo test --test pipeline_lifecycle_test passes (weak refs verify full finalization)
• Tested locally: 58 flows started+stopped, zero leaked elements in log
• Deployed to production: flow restart, socket count stable (no growth), zero ERROR in logs
• All block-level closures audited — all use WeakRef or safe types (no strong GStreamer captures)

🤖 Generated with Claude Code