[Afloat16]

DefaultStrategy updates screen radii through repeated advanced-index assignments. When one Gaussian appears in several cameras, a later smaller observation can overwrite the larger radius, making splitting/pruning depend on camera order.

Reduce observations with scatter_reduce_(amax, include_self=True), preserving both the previous history and the maximum radius in the current packed or unpacked batch. Gradient sums and visibility counts retain their existing accumulation.

Validation on Python 3.12.14 / PyTorch 2.8.0+cpu, with native full-package imports and the repository pytest configuration:

• Unchanged main: 10 failed / 2 passed / 12 CUDA skips; this branch: 12 passed / 12 CUDA skips.
• Existing strategy tests plus regressions: 13 passed / 14 CUDA skips.
• Tests cover camera order, packed/unpacked observations, retained history, invisible Gaussians and disabled screen refinement. Actual split/prune operations resize real Adam moments, preserve optimizer references and remain usable for a subsequent optimization step.
• Project-pinned Black 22.3.0 changed-file formatting, critical Ruff checks, Python compilation and git diff --check pass.

CUDA reduction execution, renderer integration, full training and image-quality/performance benchmarks were not run.