Neither built. Both are consumers of proc-nodes/src/shared, and both were last touched 08-31, before the mip-pyramid work landed in dc-terrain 332df1c on 09-01 — so the instruments were a day behind the thing they exist to measure, and the gates could not run at all. erosion_render.cpp — 14 undeclared-identifier errors, all one cause. The consumer's job changed from "define hblk_height_m" to "declare the storage": erosion_heightmap_sampler_glsl() now emits hblk_sample() and heightmap_gn calls that. So the flat bilinear hblk_height_m is DELETED rather than repaired — nothing called it, and its body still indexed hblk_data[] after the buffer was renamed to custom_data, which is what produced the errors. Added in its place: erosion_heightmap_levels_glsl() for HBLK_LEVELS/OFF/SIDE, a voxel_scale const (the wrapper emits M_VOXEL, a different name, and emits it later, so it cannot be aliased), and the WHOLE pyramid on the SSBO instead of level 0. That last one is #55 section 2.1: with level 0 only, HBLK_OFF[1] points one float past the end of the buffer. The upload sizes itself from the vector, so the count/size pair that produced the 7.9 TB allocation cannot desync here. erode_heightmap.cpp — one argument behind erosion_node_eval_full, which gained max_octaves (the fractional octave limit the engine sets per chunk from the voxel size). Passes 999 = no taper, which is what the call did before the parameter existed. This is the binary the lookdev's erosion panel shells out to, so the panel was dead too. Both verified, not assumed: erosion_render --verify matches the parity oracle at 0.0042 m worst (unchanged, and that path takes no heightmap, so it is a control on the edit); erode_heightmap --check identity 0.000e+00. gate_heightmap then passes for the first time — identity 0.005 m, control discriminates at 1078 m, worst seam z 3.48.