Resolve thread-group leader in dup_fd_from_pid for pre-6.9 kernels

[congwang-mk]

dup_fd_from_pid duplicates a client socket into the supervisor via pidfd_open + pidfd_getfd for on-behalf bind/connect/sendmsg/getsockname. When the triggering task is a non-leader thread, it relied on pidfd_open(tid, PIDFD_THREAD), which requires Linux 6.9+; on older kernels that returns EINVAL, so the on-behalf operation fails (e.g. a multithreaded client's bind returns EINVAL).

This resolves the thread-group leader via Tgid in /proc/<tid>/status and opens the leader's pidfd instead. File descriptors are shared process-wide, so the leader's pidfd can duplicate any thread's fd — on any kernel with pidfd_getfd (>= 5.6). The triggering thread is frozen on the seccomp notification, so its Tgid cannot race with pid reuse.

Verified on a 6.7.9 host (riscv64): all test_port_remap tests pass, including the multithreaded loopback / loopback_under_conflict cases that exercise this path. Architecture-independent.

🤖 Generated with Claude Code