Qualify read-only node06 P2P preflight

[nessie993]

Summary

• use node06-compatible docker top -eo pid
• ignore NVIDIA 595's non-adjacency GPU NUMA ID topology header field
• lock both real host formats with regression tests
• record the immutable tool receipt and successful read-only node06 preflight

Evidence

• 19 focused P2P tests; 100 full Python tests
• read-only node06 preflight completed in 4.0s
• exact r34, process ownership, NUMA-1 GPUs 4–7, directed NODE and P2P r/w passed
• 1,830MiB free on each target GPU
• LB/A/B IDs, starts, images, and zero restarts unchanged; LB stayed 2/2

No LB recreate, tool container, CUDA context, or benchmark ran.