nx_perf_probe.nx

nx_perf_probe.nx -- substrate performance counter primitive. Cardinal: feedback-systemic-injection-everything-zero-runtime-overhead. Same three-layered pattern as nx_bit_provenance / nx_race_telemetry: (1) Born-into-build: callers use nx_perf_tick_start / tick_end (or the wrap_call helpers below). (2) Retroactive: future self-host IR pass injects entry/exit calls driven by `// nx_perf: budget_ns=...` annotations on function definitions. (3) Zero runtime cost when --strip perf: NX_PERF_ACTIVE compile-folded to 0 + entire machinery DCE'd. Storage: ring buffer of (function_tag, start_ns, end_ns, iter) records. No per-tick syscalls; nx_perf_now() uses sys_clock_gettime_mono once per tick. API surface: let tk: i64 = nx_perf_tick_start(NX_PERF_TAG_MATMUL) // ... work to time ... nx_perf_tick_end(tk) -- elapsed recorded nx_perf_drain(out: *i64, cap: i64) -> i64 nx_perf_n_records() -> i64 nx_perf_enable() / nx_perf_disable() / nx_perf_is_active() Each tick uses ONE clock syscall + one ring-buffer write. No startup cost. No ledger allocation until first tick.