Move top-level `import torch` into _cuda_available() helper so kronos_adapter.py can be imported in CI environments without torch. All device defaults resolved at runtime via lazy detection. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add Kronos-mini (4.1M params, AAAI 2026, MIT) as: - Option A: predicted-return alpha factor via rolling daily inference (kronos_factor_gen.py — stride=96 bars/day, ~2k inference calls) - Option B: standalone model evaluator alongside LightGBM (kronos_model_eval.py — IC / hit-rate vs actual realized returns) KronosAdapter wraps NeoQuasar/Kronos-mini + Kronos-Tokenizer-2k, auto-detects GPU, gracefully degrades if ~/Kronos repo is missing. Factor output: MultiIndex (datetime, instrument) with KronosPredReturn. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
TPTBusiness