[yeseorizi]

Summary

Fix DVL-to-ROS translation and callback lifetime issues reproduced on current main (0d19c3f).

• Gazebo's linear covariance is 3×3; ROS TwistWithCovariance is 6×6. Map values into the top-left linear block, rather than copying them contiguously into the first nine ROS entries. Apply the same mapping to individual beams.
• Bound malformed covariance input to nine entries to avoid writing past the fixed ROS array.
• Create the ROS publisher before subscribing to Gazebo transport. A shared callback gate prevents in-flight / late transport callbacks from dereferencing a destroyed bridge; unsubscribe during destruction.
• Do not replace Gazebo's SIGINT handler when initializing ROS. Ignore callbacks before configuration or after ROS context shutdown.
• Remove the per-sample debug message from the callback hot path.

Reproduction and validation

The baseline failed the covariance-layout and host-signal-handler regressions. Tests also cover empty/oversized covariance, unconfigured and shut-down contexts, and 30 create/destroy cycles while a transport publisher is active (with a connection assertion).

• Independently rebuilt this branch's ROS plugin package against the unmodified main workspace with UBSan / float-cast-overflow enabled.
• All 7 cases passed in 10 repeats (70 case executions), including 300 lifecycle iterations.
• Changed-file pre-commit checks passed.

colcon build --packages-select posim_ros_gz_plugins --cmake-args -DBUILD_TESTING=ON
ctest --test-dir build/posim_ros_gz_plugins --output-on-failure \
  -R '^test_dvl_bridge$' --repeat until-fail:10

Scope / limitations

Tested in Ubuntu 26.04 ARM64 Docker on an Apple Silicon Mac, ROS 2 Lyrical / Gazebo Jetty 10.5.0. The fix concerns translation and lifecycle, not changes to the acoustic DVL model or a claim of physical accuracy. No ASan/TSan, native Linux host or real-GPU validation is claimed. The separate ros_gz_bridge shutdown issue is not fixed by this PR.

Checklist

• Forked the repository and kept the change focused.
• Added regression tests and ran the targeted functional tests plus changed-file pre-commit.
• Signed commits with DCO and disclosed AI assistance.
• Entire repository colcon test is clean. The earlier full audit also reports pre-existing ament cpplint/uncrustify/copyright failures; no tests were disabled here.
• Remote formatting CI passed. Docker jobs are intentionally skipped for fork PRs; this is not a Docker build result.

Assisted-by: OpenAI Codex (implementation, tests and PR preparation).