The spatializer stops modeling the head and starts using one. Sources are convolved with the MIT KEMAR compact HRIRs (128 taps, 44.1 kHz, 368 measurements from -40 to +90 elevation — Gardner & Martin 1994, vendored in tools/hrir with their citation): ITD, head shadow, and the pinna's elevation notches come from the measurements, not from math that imitates them. Nearest-four bilinear interpolation over azimuth and elevation, the filter refreshed every 64 samples and crossfaded across the block so a moving source never zippers, 1/r distance applied before convolution, per-source spatialization summed at the ears, and exactly one normalization at the end. Verified against directional probes: 726 us ITD at the sides and 0 on the median plane, 14 dB ILD, and above/below separated by spectral shape alone — which is how elevation actually works. Renders 60 seconds in under three. Scores gain pinned lanes: a lane with az/el/dist is a fixed thing in the room rather than a voice on the orbit. monosine uses two — hats overhead front-left, clicks underfoot front-right — so the sweeping sine has anchored references to pass, and broadband transients (which localize far more sharply than tone) carry the vertical cue. Claude-Session: https://claude.ai/code/session_01Xju7GpzP6o2JwxtqfKqFiG