nx_vpick_gate.nx

nx_vpick_gate.nx -- the gate for nx_vpick_lib (/compare/dcc DC4: ray-pick and a transform gizmo). THE CLAIM UNDER TEST is that a click in the sovereign viewport resolves to the RIGHT triangle and a drag resolves to an UNDOABLE operation. Both halves are asserted on values a wrong implementation cannot produce, never on a return code alone -- because the three cheapest wrong pickers in the world all return plausible codes: one that always misses, one that always returns triangle 0, and one that returns whichever triangle happens to be first in the buffer. THE FIXTURE CAMERA IS THE SHIPPED CAMERA. VG_W, VG_H, VG_DIST and VG_FOVH below are literally MV_W, MV_H, DIST and FOVH from nx_meshview_wasm. If the pick maths had drifted from the renderer's convention the exact-value teeth would fail, because they are the numbers that camera actually produces. ANTI-VACUITY, named where it lives, because most of these teeth are individually satisfiable by a stub: * a picker that ALWAYS MISSES passes every miss tooth -- killed by the exact-barycentric positive control, which pins u, v, w, t and the hit point to arithmetic worked out by hand. * a picker that ALWAYS RETURNS 0 passes the positive control -- killed by the named-miss tooth. * a picker that RETURNS THE FIRST HIT IN THE BUFFER passes a single depth test -- killed by running the SAME two triangles in both buffer orders and requiring DIFFERENT answers (1 then 0). A first-wins picker answers 0 twice, a last-wins picker answers 1 twice; only genuine depth sorting answers 1 then 0. * a degenerate-triangle guard that simply skips everything passes "degenerate is never picked" -- killed by putting the degenerate FIRST and requiring the valid triangle behind it to be found WITH its correct barycentrics. * a gizmo that returns ZERO is undoable, replayable and axis-constrained -- killed by comparing its magnitude against a closed form derived a DIFFERENT way (view half-width at the pivot plane), which also kills a gizmo that returns the raw pixel delta. * an edit stack that records nothing has a stable digest -- killed by requiring the digest to CHANGE and by reading the moved vertex back component by component. * a BENCH whose loop the compiler deleted times as FREE and reads as a very fast picker -- killed by accumulating the picked index and requiring the sum to equal the repetition count times what a single pick returns, a number no deleted loop can produce. * a LINEARITY verdict taken at ONE triangle count is not a verdict at all, because a quadratic curve and a linear one are the same number at a single N -- killed by four rungs spanning 64x, and by a bar derived from the two hypotheses being separated rather than tuned until the result passed. Every guard in the library is a gv_bite: it must fire on the bad input AND stay silent on the good one. THE SECOND HALF OF THE RUNG, ADDED 2026-08-25. A pick that is CORRECT is still not an interaction claim until it carries a COST at a stated triangle count -- dcc.plan's unit line says exactly that, and the rung's done-rule was half unmet while this file proved only correctness. Section 9 times vp_pick_screen over procedurally generated grids at four triangle counts, publishes nanoseconds per operation with its