nx_ctxtop_lib.nx

nx_ctxtop_lib.nx -- PURE core of nx_ctxtop: per-process context-switch ATTRIBUTION. WHY: nx_procchurn proved the box takes 81k-93k context switches/sec in EVERY sample across 45 minutes (1.6-1.9x the RED line) -- the one CONSTANT in the perf lane. But a box-level rate names no culprit, and the fork-rate hypothesis was DISPROVEN as the driver (seq1340: forks swing 15->179->10 while ctxsw stays flat, so the switches are NOT mostly fork/exec). /proc/<pid>/status carries voluntary_ctxt_switches and nonvoluntary_ctxt_switches PER PROCESS, so the producer is directly measurable instead of guessed. LAW (rising-debt-means-find-the-producer): a rate with no owner is a symptom, not a diagnosis. ★THE VOLUNTARY/NONVOLUNTARY SPLIT IS THE DIAGNOSIS, not a detail: voluntary = the process BLOCKED on its own (read/poll/sleep) -- an I/O or poll-loop design, fixable by batching, longer polls, or event-driven waits. nonvoluntary = the SCHEDULER preempted it -- CPU contention / too many runnable threads, fixed by reducing concurrency, not by batching. The same total means opposite remedies, so a tool reporting only the sum would send the fix the wrong way. This lib keeps them separate all the way to the output. Rule 15: rate arithmetic is REUSED from nx_procchurn_lib (pc_rate), generic IO/print from nx_resmon_lib -- no new copies of the emit family (D001/seq389: 9495 duplicate bodies). license_tier: ORIGINAL Read-only. No hw writes (Rule 26).