argv-safety: register children at fork time

[congwang-mk]

Summary

• When any consumer can inspect execve argv from child memory — policy_fn, or an extra handler bound to execve/execveat that can call read_child_mem — the supervisor now arms one-shot ptrace fork-event tracking (PTRACE_O_TRACE{FORK,VFORK,CLONE}) on every fork-class notification. New children are registered in ProcessIndex before they can run user code, so the execve argv-safety freeze can enumerate every tracked TGID and PTRACE_SEIZE every thread that could mutate argv between supervisor inspection and the kernel's post-Continue re-read.
• Default policies are unaffected: bare fork(2) only enters the BPF notif filter when policy_fn is set or an extra handler is bound to exec, so the COW map-reduce hot path keeps bypassing the supervisor.
• ProcessCreationTrace is an RAII guard — panics or early returns can't leak ptrace attachments. The blocking waitpid calls run on tokio::task::spawn_blocking so a stretched wait can't stall a notification worker.

Test plan

• cargo test --workspace — 443/443 passing (was 441; +2 new tests in crates/sandlock-core/src/resource.rs)
• pytest python/tests/ — 226/226 passing
• process_creation_tracking_predicates_follow_argv_safety_gate locks in the gating across clone/clone3/fork/vfork and the CLONE_THREAD bypass
• process_creation_tracking_registers_child_before_user_code_runs validates the central safety property end-to-end via a MAP_SHARED flag page: a raw-fork child is observed to be registered while still ptrace-stopped (before its first user-mode instruction). Skips gracefully if YAMA ptrace_scope denies the seize.
• x86_64-only integration test (raw fork(2) syscall); other arches skip via #[cfg] — intentional

🤖 Generated with Claude Code