[Afloat16]

torch.cdist switches to a matrix-multiply distance identity for larger point sets. Subtracting large squared world coordinates loses the small relative distances: translating 32 unit-spaced points by 10,000 can change initialized neighbor scales from 1 to 1e-7.

Use direct Euclidean differences in each existing chunk, mask each query's actual self index and select its k nearest other points. Chunking and the log-scale/epsilon contract remain unchanged.

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

• New regressions on unchanged main: 7 failed / 8 passed; this branch: 15 passed.
• Existing multiframe/init tests plus regressions: 25 passed / 1 CUDA skip.
• Independent relative-distance references cover float32/float64 translations, multiple k/chunk sizes, duplicate coordinates and coordinate gradients.
• Project-pinned Black 22.3.0 changed-file formatting, critical Ruff checks, Python compilation and git diff --check pass.

The direct distance mode trades throughput for numerical stability. A bounded single-thread CPU microbenchmark (k=3, chunk=256, five measured repetitions) gave median 2.01→2.73 ms for 512 points and 21.28→34.13 ms for 2,048 points. These are local observations, not a GPU or complete-training performance estimate. CUDA, full reconstruction/rendering and large-scene memory benchmarks were not run.