nx_perf_lib.nx

nx_perf_lib.nx -- FRAME-BUDGET TELEMETRY, the AAA way, as ONE sovereign ruler (GE32 + GE15, 2026-09-02). Operator: "evaluate at minimum the way AAA teams do -- every frame is a budget that cannot be overspent; frame time and frametime CONSISTENCY, spikes, p50/p95/p99". The estate measured all of that in page JavaScript (?perf=1: p50/p95/p99/jank/worst) and NOTHING read it (gameengine GE15). This lib moves the arithmetic into NishiLang so the ENGINE (wasm), the served HUD and the pacing referee gate all compute the same numbers from the same window -- one ruler, three consumers -- and the page is left with a door call and a display line, which is the interop-only law. WASM-COMPATIBLE BY CONSTRUCTION: no allocation, no syscalls, no floating point. The caller hands in a region of PF_WORDS i64 (an engine arena offset, or a native mmap) and every function is a pure operation over it. Integer milliseconds throughout; the page rounds the rAF interval before it crosses. WINDOW vs CUMULATIVE, deliberately both: a ring buffer reports the RECENT, not the WORST (estate law), so the header carries cumulative counters (frames ever, worst ever, over-budget ever) beside the PF_N-frame window the percentiles are computed over. Percentiles use the nearest-rank method over a 0..PF_CLAMP_MS histogram (rank = ceil(permil x n / 1000)). pf_push preserves legacy capped input; pf_push_raw retains stalls beyond the histogram and resolves their ranks from the raw ring. An EMPTY window answers PF_UNOBSERVED (-1) to every statistic -- a check that reads 0 frames as 0 ms would pass on the empty set. COMPOSES nx_frame_budget (the estate's per-frame budget ruler, microseconds): the budget in ms for a target fps is fb_budget_us(fps)/1000 and fits/headroom are fb_fits/fb_headroom_us -- not re-derived here. license_tier: ORIGINAL No hw writes (Rule 26).