nx_perf_pct_gate.nx

nx_perf_pct_gate.nx -- KNOWN-ANSWER GATE for the frame-time percentile arithmetic the served page publishes (gameengine GE50, symbol ft_percentile_kat). WHY THIS GATE EXISTS, from a mirrored primary source rather than a hunch: the industry's reference frame-time instrument, Intel PresentMon, records in its own v2.5.0 release notes (2026-06-29) that its percentile calculation was WRONG and that 99 percent was effectively reporting MAX. A percentile that quietly degenerates to the maximum turns every p99 claim into a worst-frame claim, and nobody can tell from the number. Our p50 p95 p99 are computed by nx_perf_lib (pf_pct) inside the wasm engine and beaconed by the page, and until this file they had NEVER been checked against a sequence whose percentiles are known by construction. The lib's own header carries the claim "p50 of 1..100 is 50, p99 is 99, p100 is 100" as PROSE; this gate makes it a TOOTH. METHOD UNDER TEST (nx_perf_lib.nx): nearest-rank over a 0..PF_CLAMP_MS integer-ms histogram, rank = ceil(permil x n / 1000), clamped to [1, n]; an empty window answers PF_UNOBSERVED. Every expected value below is DERIVED from that definition and printed with gv_check_eq, so the number tested and the number published cannot drift apart. The decisive tooth is the neg-control: a window of 99 identical frames and ONE outlier has p99 == the common value and worst == the outlier; an implementation with PresentMon's defect returns the outlier for p99 and FAILS here. In-process (imports the lib): a gate that forked the wasm would prove the module, this proves the ruler. license_tier: ORIGINAL No hw writes (Rule 26).