Stabilize encounter frames for collinear relative states

[sylvesterkaczmarek]

Summary

Use a double cross product and a scale-aware degeneracy test in encounter_frame(). Subtractive projection can normalize roundoff parallel to the relative velocity, producing a non-orthonormal basis and making relative_encounter_covariance() reject a valid collinear encounter. The existing deterministic fallback is preserved.

Repository Policy Checklist

• All review gates in AGENTS.md are complete. Machine-generated code and tests; executed checks and remaining gates are recorded below.
• Changes preserve the existing file layout.
• Package-code changes include tests under tests/.
• No new workflow or demo is needed; this corrects existing numerical behavior.
• No generated outputs, data, images, or binary artifacts are committed.
• No persistent data is added.

Validation

• Full pytest suite is green; see results below.
• Full demo gallery run (not run).
• Repository policy check inspects both changed paths.
• No existing test, tolerance, workflow, or policy check was weakened.
• Every numerical result below came from an executed command.

Python 3.12, macOS arm64; published SSAPy 1.1.10 and SSAPy-Data 0.1.5. Ruff, compilation, diff check, warning-as-error Sphinx build, wheel/sdist builds, and twine check passed.

New regression cases: 12 failed and 3 passed before the fix; all 15 pass afterward. The focused conjunction suite passes 27 tests. Coverage includes covariance/probability integration and 100 ordinary-state comparisons. Full suite: 986 passed, 20 skipped, 4 failed; 43 subtests passed.

The same four full-suite failures reproduce on unchanged upstream 262b43c: orbital-maneuver demo, sibling data-directory resolution, orbit-STM finite differences, and multi-segment 6-DoF targeting. They are outside this change. The existing failures remain unresolved; upstream CI must still be checked.