nx_vcodec_nf_train.nx

nx_vcodec_nf_train.nx -- TRAINER for the learned decoder-side restoration filter (NEURAL track rung 1). Sovereign, integer-only, no floats anywhere. For every (codec-mode in {legacy tf=1, t8-rich}, qp in {12,20,32}) it runs OUR real codec (key + P chain, in-loop deblock -- exactly the live display path) over the REAL frames 0..7 of bframe_test_decoded.yuv (576x1024 luma; frames 8..11 stay HELD OUT for the gate), accumulates per (qp-band, gradient-class) integer normal equations over (3x3 deblocked-recon context -> source pixel) pairs, then solves each class's 3x3+bias Q12 filter: step 1 CLOSED FORM: the center-vs-3x3-mean Wiener blend a* = S_cm_s / S_cm2 (scale-free integer) step 2 REFINE: damped diagonally-preconditioned integer gradient descent on the full 10-tap least squares (J tracked from the moments; a step that raises J is reverted + step halved), starting AT the blend -> the result can only match or beat it on the training distribution. Classes with too few pairs keep the IDENTITY filter (never invent weights from noise). Emits the table as GENERATED SOURCE runtime/nx_vcodec_nf_table.nx (vc_nf_load) -- generators, not hand-authored data. license_tier: ORIGINAL