ptqtp-spirals — post-training trit-plane quantization, self-verifying

Two-spirals PTQTP acceptance demo (main.zig): train a float tanh MLP with AdamW (2 -> [dense f32] -> 256 -> tanh -> [256x256] -> tanh -> [head 256x2]), then post-training-quantize the two packable layers to dual trit-planes (fucina.ptqtp, arXiv:2509.16989) and measure what survives — no retraining, no calibration data, weights only. See docs/PTQTP.md for the method.

The report compares, on the same raw-kernel eval harness:

variant	what it is
float	the trained dense weights (the ceiling)
absmean-b1.58	blind round-clip, one plane (the zero-optimization floor)
ptqtp-k1	one plane, ridge scales + 3-way search
ptqtp-k2	the paper's dual planes, 9-way search

each ternary variant on both forwards: the exact mul-free f32 path (isolates weight-approximation error) and the deployed int8 path (Q8_K activations x packed crumbs — adds activation quantization). It also prints packed reconstruction errors, plane sparsity, and the reference-path G=128 vs G=256 fidelity delta.

Self-verifying: exits nonzero unless the float model reaches 0.99 training accuracy (else inconclusive) and the dual-plane int8-path accuracy reaches --target (default 0.95; chance 0.50).

zig build ptqtp-spirals -Doptimize=ReleaseFast
# all knobs:
zig build ptqtp-spirals -Doptimize=ReleaseFast -- [--steps N] [--seed N] [--target F] [--lr F]

Defaults: --steps 3000 (stops early at 100% train accuracy), --seed 42, --target 0.95, --lr 0.02.

A healthy run ends with a PASS: line (float 1.000 -> ptqtp-k2 holding accuracy on the deployed int8 path); the measured reference table for this exact demo is in docs/PTQTP.md under Measured — single ternary planes collapse, the dual decomposition holds.

Shared knobs

The ReleaseFast/-Dcpu build discipline and global thread/BLAS knobs are shared machinery — see docs/RUNNING-MODELS.md. No model weights needed; for PTQTP on a real LLM see ../ptqtp_qwen3.