RFD 1126: Unroll the descent, do not replace it
Problem
RFD 1122’s chain has six stages: three feed-forward graphs, and three descent loops that an inference accelerator cannot compile.
The obstacle is control flow rather than precision. lbfgs_polish.py uses a strong-Wolfe line search, a loop whose trip count depends on values, and Loop and If are what an edge compiler refuses. Its float64 follows from that same file’s tolerance_grad=1e-10, an offline target rather than an inference one.
An earlier draft answered by replacing the fit with a learned student, which gives up the property the parametric model exists for: soma_referee.py states that a parametric model cannot represent an impossible skeleton, and nothing in a regressor’s architecture forbids one.
Decision
Unroll the descent into a fixed number of steps, keeping the forward we own.
Clamp every step, with pairwise kusudamas concatenated. No kusudama carries two or more cones: three equidistant cones sum to zero. A clamp is Clip, so the unrolled graph still cannot emit an unreachable pose.
Warm-start from the previous frame, so the step count stays small.
Compile the backbone with num_windows=1. Measured: it parses and num_windows=2 does not. It costs 1.35x wall-clock and needs retraining.
Give the face camera the resolution and starve the body cameras. One device holds four streams at 120 fps only if they differ.
Report the residual as a percentage of stature, as soma_referee does.
See DETAILS.md for the rig budget, the unrolled cost, and what is unmeasured.