nx_lowrank_kv_gate.nx

nx_lowrank_kv_gate.nx -- MEASURED Pareto gate: low-rank KV-cache compression via Frequent Directions. module: nishi-core.ai.lowrank_kv capability: VRAM_REDUCTION (the quality-UP / VRAM-DOWN lever, NOT quant) THESIS (operator 2026-06-16): innovate where math REDUCES VRAM while PRESERVING quality -- the opposite of quantization (which shrinks quality AND size). The KV cache is up to ~70% of inference VRAM at long context; attention logits depend on K^T K, so if a sketch B keeps ||K^T K - B^T B|| small we keep attention fidelity while storing far fewer bytes. Frequent Directions (Liberty 2013, our sovereign sketch_freq_directions.nx) gives exactly that with a PROVABLE bound: ||A^T A - B^T B||_F <= ||A-A_k||_F^2/(l-k). THIS GATE MEASURES the Pareto by RUNNING (no self-grading): PERSISTENT KV bytes: full = n*d*8 vs sketch = l*d*8 (CONSTANT in n -- the long-context win) QUALITY (covariance fidelity): rel error ‰ of K^T K vs B^T B (lower = attention better preserved) T1 sketch bytes are CONSTANT as context n grows (16 -> 64): VRAM does not explode with O(n). T2 full KV bytes GROW with n (the baseline we beat). T3 low-rank data: sketch fidelity error << full-rank data error (the metric is REAL, not vacuous -- FD can compress genuine low-rank structure but CANNOT fake-compress full-rank => neg-control). T4 low-rank covariance error is below an absolute quality bar (quality PRESERVED, measured). HONEST SCOPE: REFERENCE-SCALE (FD v1 caps d<=8, l<=4; real d=4096 needs FD v2 Jacobi/QR per its own roadmap note). This proves the PRINCIPLE + the provable bound, sovereignly. Reuses FD (DRY). license_tier: ORIGINAL