Add calibration-free engine_kv module for inference-engine integration (C15a)

[jagmarques]

Adds nexusquant/integrations/engine_kv.py: a clean KV-quant/dequant API for vLLM-style engines, plus a CPU unit test and GPU bench scaffold.

What it does

• quantize_kv(k, v, k_bits=3, v_bits=2) returns a PackedKV (int8 codes + fp16 per-head scales). Hadamard rotation first, then E8 nearest-point, then pack as (lattice_point * 2).int8. No calibration, no fit step, no offline data.
• dequantize_kv(packed) reconstructs fp16 K/V via (code * 0.5) * scale — scalar multiply only, no codebook lookup.
• bpe_accounting(packed) reports logical bpe (K3V2 = 2.625 bpe, 6.1x vs fp16) and int8 storage bpe (8.125 bpe, ~2x).

CPU unit test (experiments/kaggle/nq_engine_kv/test_engine_kv.py, exit 0 verified):

• (a) K3 round-trip relative error 0.1918 (bounds [0.01, 0.35]); V2 error 0.3864 (bounds [0.01, 0.60]).
• (b) bpe_logical = 2.6250 (expected 2.6250, tolerance 0.05).
• (c) 87.2% of k_codes differ from trivial round(k_fp32) — non-vacuous.

GPU bench scaffold queued in experiments/kaggle/nq_engine_kv/ (kernel-metadata.json pinned to same T4 docker SHA). Not pushed to Kaggle (quota exhausted).

Proof of work:

python3.11 experiments/kaggle/nq_engine_kv/test_engine_kv.py
[PASS] (a) K3 rel error: 0.1918 (expected ~0.19, bounds [0.01, 0.35])
[PASS] (a) V2 rel error: 0.3864 (expected ~0.39, bounds [0.01, 0.6])
[PASS] (b) bpe_logical: 2.6250 (expected 2.6250)
       ratio_logical: 6.10x  bpe_int8: 8.1250
[PASS] (c) non-vacuous: 87.2% codes differ from trivial identity
       max element diff after round-trip: 0.9512 (>1e-3 required)
All 3 assertions passed. exit 0.
EXIT_CODE: 0

secret-scan: SCANNER: gitleaks - clean (exit 0)
diff stat: 4 files changed, 526 insertions(+)