macOS CPU build + toroid phantom-collision fix (#467)

[treegardel]

Type of Change

• Bug fix (non breaking change, fixing an issue)
• New feature
• Breaking change (fix or feature that causes existing functionality to not work as expected)
• Documentation / Wiki update

Description

This PR bundles two related sets of fixes that came up while getting the macOS CPU build green.

1. macOS CPU build pipeline

• Set macOS deployment target to 13.3 and pass it consistently to CMake.
• Disable rayx-ui and CUDA only for macOS CI.
• Update CMake Apple logic so CMAKE_OSX_DEPLOYMENT_TARGET is only set as fallback (does not override workflow value).
• Add AppleClang to platform detection so RAYX_API exports symbols correctly on macOS.
• Implement macOS getExecutablePath() / getFullPath() in Locate.cpp (was static_assert(false)).
• Drop the spurious baseDir segment from the macOS resource lookup so Data/PALIK, Data/NFF, Data/CROMER, Data/MOLEC are found next to the executable (matches Linux behavior). Without this, all material tables came back empty and getRefractiveIndex returned (-1, -1).
• Fix formatAsVec redefinition on Linux (unsigned long == RandCounter).

2. Toroid phantom collisions — closes #467

The Newton solver in getToroidCollision converges to within NEW_TOLERANCE = 1e-4 mm. After a hit, the ray is nudged forward by COLLISION_EPSILON = 1e-6 mm — much less than the Newton residual. The forward check tested only the sign:

if (dot(rayToHitpoint, rayDirection) <= 0.0) return std::nullopt;

A solution within the residual on the wrong side of the new origin still passed, so each toroid bounce was counted ~4× in non-sequential tracing.

Fix: reject hits inside Newton's own uncertainty by checking parametric forward distance, local to the toroid solver:

const double t = dot(rayToHitpoint, rayDirection) / dot(rayDirection, rayDirection);
if (t <= NEW_TOLERANCE) return std::nullopt;

COLLISION_EPSILON stays at 1e-6 mm, all other surfaces keep their precision.

Tests added:

• Integration test toroid_iteration (grazing toroid vs. RAY-UI reference, regenerated at proper 1e-4 tolerance — old reference was at 1e-1 and hid the agreement).
• Unit tests on getToroidCollision: forward & backward grazing hits at z ≈ ±600 mm walking a tolerance series from 1e-2 down to 1e-6 mm (1 nm), plus a phantom-rejection regression test. Newton's quadratic convergence delivers sub-nm precision in practice.

Bonus: getCubicCollision had 29 pow(float(expr), N) calls truncating doubles to single precision on every Newton iteration — capping cubic-surface accuracy at ~1e-7 relative regardless of tolerance. Removed; all tests still pass.

This is not a breaking change. Linux/Windows workflows are not functionally changed by this PR.

Fixes #467

---

✅ Pre-Merge Checklist

• Code follows the project's coding standards
• Unit tests for new functionality are added and pass
• All existing tests pass
• Resolved TODO Comments (prefer new issues instead)
• Documentation, if applicable, including:
  • Doxygen comments for any new rayx-core API functions (N/A)
  • Helpful inline comments where needed for clarity
  • Wiki pages, e.g. updated build instructions, new Element etc. (N/A)
• Commits:
  • Use clear and readable commit messages
  • Squash and rebase onto master if individual commits don't add value
  • Ensure linear commit history (required by master)

---

ℹ️ Supersedes #465 — der ursprüngliche PR wurde versehentlich gemergt und der Merge anschließend per Reset aus master entfernt. Da ein gemergter PR auf GitHub nicht wieder geöffnet werden kann, ersetzt dieser PR ihn inhaltlich unverändert.