fix(core): Do not raise floating point exceptions during create/init

[hgopalan]

Fixes #404.

MoorDyn_Init() (and in some configurations MoorDyn_Create()) raised FE_OVERFLOW, FE_DIVBYZERO or FE_INVALID on intermediate values that are discarded afterwards. The results are correct, but a host program that traps floating-point exceptions (feenableexcept() on Linux, FPCR trap bits on macOS arm64) gets killed. While checking the maintainer's question in #404 (whether anything else trips once the CFL overflow is patched), I ran the whole test suite with the traps enabled and found three more sites. This PR fixes all of them.

Changes

Site	Exception	Fix
CFL::cfl2dt(cfl, v): cfl * max() / v for points, rods and bodies (the case reported in #404)	overflow	return the max() sentinel unchanged
CFL::cfl2dt(cfl, v) with v == 0 in StationaryScheme::Step()	div-by-zero	return max()
NatFreqCFL::cfl2dt(cfl) with cfl = max(), which MoorDyn::Init() sets when dtM is prescribed, for lines whose segment natural period is > 1 s	overflow	return the max() sentinel unchanged
Rod::setup(): i / (real)N with N = 0 for zero-length rods	invalid (0/0)	use f = 0
Catenary() Newton-Raphson iterations (weightless or buoyant lines, log/sqrt of negatives, x/0) called from Line::initialize()	invalid, div-by-zero, overflow	hold the FP environment around the call (feholdexcept/fesetenv) and discard the flags; the solver already reports the failure and the line falls back to the linear profile

In every case the old value was inf/nan and was already dropped by std::min or by the caller, so results do not change. The catenary change leaves the solver itself untouched. If you'd rather guard each operation inside Catenary(), I can do that instead, but it is a much more invasive change to that routine.

tests/fpe.cpp adds a portable test (no traps needed) that clears the flags, then runs create, init and 10 steps on 5 inputs covering the sites above, and requires fetestexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW) == 0.

Testing

macOS 26 arm64, AppleClang, RelWithDebInfo, PYTHON_WRAPPER=OFF:

• ctest: 32/32 pass.

• ctest with the traps enabled in every test process (a preloaded library sets the FPCR trap bits):

  trap	dev	this PR
  overflow	19 / 31 tests killed	0
  div-by-zero	9 / 31 tests killed (with the CFL overflow fixed)	0
  invalid	12 / 31 tests killed (with the CFL overflow fixed)	0
  all three	—	0 (32/32 pass)

• tests/fpe.cpp on dev: 5/5 cases fail; with this PR: 5/5 pass.

• The *.out files written by the test suite (76 files) are byte-identical between dev and this branch.

• The reproducer from MoorDyn_Init raises FE_OVERFLOW in CFL::cfl2dt (stationary IC solver); fatal when the host traps floating-point exceptions #404 (line between two fixed points in air, external kinematics) runs Init plus 60 s of simulated time (6000 steps of 0.01 s with a gusty cross-flow) with all three traps enabled. It raises no flags and gives the same fairlead tension as dev.

🤖 Generated with Claude Code