fix: harden event sends on shutdown, debounce connection-triggered re-elections

[caesarniseko-coder]

Summary

• Shutdown hardening: event sends in DownloadCoordinator and Worker._forward_info now tolerate closed/broken channels instead of crashing, so components exit cleanly during shutdown
• Re-election resilience: apply() skips out-of-order indexed events with a warning instead of asserting, so replayed events after a master re-election no longer crash the node
• Topology correctness: thunderbolt bridge cycles are recomputed when network interfaces or thunderbolt connections change, not only when bridge status changes
• Election debounce: after a connection-triggered campaign, further connection messages are absorbed for 25s and coalesced into at most one follow-up round. This stops topology-formation bursts from causing perpetual re-elections, while never dropping a master-disconnect (a follow-up round runs if anything arrived) and never delaying shutdown (a closed channel ends the wait immediately)

Test plan

• uv run pytest — full suite green (370 passed); election tests 14/14, suite runtime dropped from ~52s to ~2s since shutdown no longer stalls
• uv run basedpyright — 0 errors
• uv run ruff check — clean
• Manual: node boots, API responds, SIGINT shuts down cleanly with no errors logged

🤖 Generated with Claude Code