nx_lowrank_rank_plan.nx

nx_lowrank_rank_plan.nx -- DATA-DRIVEN per-tensor/per-layer KV rank sizing. Bridges M2 ("the optimal bottleneck rank MATCHES the data's true rank") + M4c-1 (the REAL Qwen2.5-0.5B KV spectrum is per-tensor: K steep, V shallow) into an ACTIONABLE rule -- with NO magic ranks (rule #11): the energy threshold is the ONLY knob; each (layer,tensor) rank is DERIVED from its measured cumulative-covariance-energy spectrum. rank(layer,tensor) = min r such that cumulative_energy(top-r) >= threshold Pure integer math on permille energy -> NO slow LLM forward (operator: "sized from the already-measured spectrum, no slow forward needed"). Validated on controlled spectra; applied to M4c-1's real anchors. 100% sovereign (nx_cc->nxasm, no gcc). license_tier: ORIGINAL ⚠ SUPERSEDED 2026-06-17: the real-data figures here (K=1/V=5 @90% -> 34x) came from a DEGENERATE-context KV spectrum (M4c-1 anchors). The definitive seq=64 measurement (nx_lowrank_kv_real M4d) shows real KV is only ~2-3x compressible at 90% energy (avg rank K=27/V=37 of 64). The data-driven RULE below is VALID; the MAGNITUDES are wrong. WEIGHTS are heterogeneous too (nx_lowrank_weight_spectrum): attn_q is low-rank but SMALL, the big ffn weights (most params) are NOT low-rank -- so weight-SVD also saves little OVERALL. BOTH low-rank VRAM levers are MODEST on this real model. Kept as the RULE/method demonstration only.