nx_enginelab.nx

nx_enginelab.nx -- THE DRIVER: attach the sovereign engine instrument to a REAL renderer. WHAT WAS MISSING AND WHY IT MATTERED. nx_enginelab_lib (the spine), nx_enginelab_store (EL1 persistence), nx_enginelab_golden (the O3DE class) and nx_enginelab_analysis (replay, shader binding, sampling, call tree) are roughly 69 KB of gate-proven library. The gate that proves them says so in its own verdict: every tooth asserts an arithmetic property "on fixtures whose answers are known by construction". That is a proof about the INSTRUMENT and it is worth having -- but until this file existed NOTHING drove the instrument over a real frame, so the estate owned a profiler that had never profiled anything. A measurement capability with no driver is BUILT-AND-UNWIRED, and a green gate makes that invisible: the gate passes exactly as well with zero real subjects as with a thousand. THE SUBJECT, AND WHY THIS ONE. nx_wasm_craft is the estate real renderer -- a per-pixel integer DDA raycast over a voxel world plus a skinned-mesh character composite -- and it is the ONE render path reachable natively with no browser (nx_world_snap already drives it that way, which is the standing evidence that init_impl_v / tick_impl / render_impl compose outside a page). Attaching here therefore needs no GUI, no capture harness and no second copy of the engine. WHAT THIS DRIVER MEASURES, AND WHAT IT REFUSES TO. It owns the call sites it wraps: the arena allocation, init, and per frame a tick zone and a render zone. Everything derived from those is a real observation of a real frame. It does NOT own the inside of render_impl, so: - THE DRAW CLASS IS UNMEASURABLE HERE AND SAYS SO BY NAME. The engine emits no draw stream -- there is no el_draw call anywhere in nx_wasm_craft -- so draws, redundant binds, pipeline state hashes, capture diff and replay divergence have NO SUBJECT. This driver reports draws=0 with the reason attached and the exact emission site the engine lane owes. It does NOT reconstruct the engine visibility test from outside: mob_draw culls on 200 < vz < 40*256, and copying that predicate here would be a duplicate ruler that drifts silently the day the engine changes it. A NAMED ABSENCE IS A BUILD CONTRACT; A RECONSTRUCTED NUMBER IS A LIE WITH A DENOMINATOR. - THE BOUND VERDICT ABSTAINS. This renderer is a CPU integer raycaster; nothing is submitted to a GPU queue, so el_bound_verdict correctly returns UNKNOWN. Manufacturing a GPU signal to make the verdict decide would be the acquit-on-no-evidence defect the spine exists to prevent. What it CAN answer, and nothing else in this estate could before: how long a real frame costs, how that cost splits between simulation and rasterisation, what the cost per RAY is, whether the pacing stutters, and whether two independent runs of the same seed produce the same pixels. RAYS ARE THE UNIT OF WORK, AND THE COUNT IS ENGINE-OWNED. wc_rw/wc_rh are the engine own single owner of the written extent (its source says so), so rays-per-frame is READ from the engine rather than recomputed here. ns-per-ray is the figure that survives a resolution change; a bare frame time does not.