[codex] Fix latest main test failures

[atimics]

Summary

Fixes the latest main test failures observed after rebasing onto origin/main.

• Makes ScalarEncoder.decode resolve near-ties with the scalar cosine path that the cached decode is meant to match.
• Stabilizes anisotropic splat early-stop and coarse-to-fine densify defaults under the current NumPy/BLAS runtime.
• Widens the private MIS generation probe beam and updates the recorded measurement so the verifier setup is real again while the durable no-op claim remains exact.

Root Cause

The failures were mostly runtime-sensitive numeric boundary cases:

• Cached scalar decode and the old scalar loop summed in different orders, flipping a midpoint tie.
• The anisotropic splat optimizer landed exactly on an early-stop boundary and the densify schedule used Adam settings that were too aggressive on this runtime.
• The MIS probe's actual invariant, balance == verifier, still held exactly, but the narrow beam made the setup assertion brittle.

Validation

• Targeted failing subset: 7 passed
• Related files: 26 passed
• Full suite, unsandboxed for localhost/process-pool tests: 3761 passed, 20 skipped