nx_hget_leak_probe.nx

nx_hget_leak_probe.nx -- MEASURES the per-call allocation leak in ss_hget (2026-07-30). WHAT IT PROVES: ss_hget used to sys_mmap TWO scratch buffers PER INVOCATION and never free them. sys_mmap is a real mmap syscall (nx_syscalls.nx:167), not a bump allocator, so the kernel rounds each to a full page => ~8 KB leaked PER CALL. ss_hget is the hottest read primitive in the tree (205 call sites; the loaders call it ONCE PER ROW), so the leak scales with ROWS SCANNED, not with stores opened -- which is why it hid behind the per-open leak everyone was looking at. THE MEASUREMENT IS THE POINT: run N hgets and read our OWN VSZ before and after. A fixed ss_hget must show delta_kb == 0 (scratch allocated once, on first call). The pre-fix code would show roughly N * 8 KB. Reuses nx_leak_check_lib (lc_vsz) rather than re-deriving /proc parsing -- rule 15. NON-VACUITY: the loop must do REAL work, so the probe seeds a store, reads a key back, and asserts the VALUE is correct. A probe that measured memory while looking up a missing key would pass while exercising almost nothing (ss_hget returns early on ns==0, before the allocations).