Replace the primary Pyramids safetensors bundle with training step 98,800, trained using simple + medium + expert data. CUDA cached-runtime validation over seeds 100..107 (128 episodes per context): ctx16 1.7158, ctx32 1.7509, and ctx64 1.5994 mean return. Keep the simple-only bundle under pyramids/simple and label the preceding ONNX exports as legacy.
Add the Pyramids 60k safetensors bundle and fixed-batch-16 ctx16/ctx32/ctx64 ONNX policies. Closed-loop validation covered 64 episodes per context over Unity seeds 100..103; ctx16 is the recommended variant at 63/64 reaches and mean return 1.7503.
Replace the legacy ONNX-only policy set with reproducible bundle-format-v2 DungeonEscape and SoccerTwos policies. DungeonEscape KV32 validation: 69/108 ego-agent and 34/36 cooperative-group successes. SoccerTwos: 33/80 wins against the release-23 stock policy. Prior ONNX release is preserved at tag pre-unity-safetensors-bundles-20260801.
Add Pyramids ONNX policy and 128-episode benchmark
