mn-fpfn30-longctx — king-line base + numpy ForecastPFN axis at 0.30

Mainnet-91 candidate (2026-07-16). Base = the byte-identical reference 10-family numpy custom_miner every top-5 entrant runs (sha 482b0c98… lineage), with the king's proven levers kept verbatim: all-4096 (min_length = max_length = 4096 = training context), de-trended bimodal excursions (tr_exc 0.4/2.5, gr_exc 0.3/1.5), chunked interleaved emission.

Added: an 11th family forecast_pfn — a pure-numpy, fully batched ForecastPFN-shaped prior (trend × decaying multi-harmonic seasonality × mean-normalised multiplicative Weibull noise + occasional level shifts and decaying spikes). The fpfn distribution held the longest testnet KOTH reigns, but on mainnet every fpfn attempt (iris999 0.55, tora 0.42) shipped the torch/pandas original and died (failed_train / untrained sentinel). This delivers the same statistical shape at numpy batch speed — the one axis the field has never successfully trained on.

Weights (sum 1.00): fpfn .30 / trend_seasonal_ar .22 / regime_shift .10 / integrated .10 / ar2 .08 / rff_gp .08 / multiplicative .05 / threshold_ar .03 / chaotic .02 / intermittent .01 / pulse_outlier .01.

Numpy-only, deterministic per (seed, series index).