nx_energy_model.nx

nx_energy_model.nx -- STATE-INDEPENDENT energy estimation, pure NishiLang. The S-class answer to "you can only measure energy when not charging": you do NOT measure the battery at all. Energy is estimated from a CHARGE-INDEPENDENT activity counter -- cycles (rdtsc/rdcycle, available on EVERY target including a PMU-less VM) or retired instructions (perf, where a PMU exists) -- times a per-unit energy coefficient: energy = activity_count * energy_per_unit activity_count does not depend on whether the node is plugged in, so this yields a real energy number on AC, on a battery-less server, on a supercomputer node, or on a discharging garden sensor -- ANY power state, ANY hardware. The coefficient is the only chip-specific part. A real joule sensor (RAPL, an INA219 on the rail, or a discharging fuel-gauge) CALIBRATES it once: energy_per_unit = real_energy_measured / activity_count_for_that_work Where no sensor is present we use a documented PRIOR (clearly labeled, not a measured claim) and refine it the moment a sensor appears. So a sensor IMPROVES the number; it is never REQUIRED to get one. Research (Cardinal #4, real sources): per-operation / per-cycle energy figures -- Horowitz, "Computing's Energy Problem (and what we can do about it)," ISSCC 2014; event-driven accounting -- Bellosa, ACM SIGOPS EW 2000; PMC power models -- Isci & Martonosi, MICRO 2003; Bircher & John, ISPASS 2007.