Unify propagator force model; cap gravity coefficient table

[ssmichael1]

Summary

Two internal refactor/perf changes to the high-precision propagator and Earth-gravity model. No behavior change — end states are bit-identical and the full Rust test suite (including the ESA SP3 GPS regression) passes unchanged.

1. Force model unified across all integrators

The RKV ydot, RODAS4 ydot_vec/jac_fn, and Gauss-Jackson accel_fn each carried a near-identical copy of the acceleration physics (gravity, Sun/Moon, solid tides, GR, SRP, drag, thrust) — ~200 lines duplicated four ways, where adding a force term meant editing every copy in lockstep.

• A single force_model() now holds the physics; each integrator closure is a thin adapter over its own state layout (Matrix<6, C> / Vector<f64, 6> / (r, v)).
• A ForceEval mode selector (Accel / AccelAndPartials / PartialsOnly) preserves each integrator's cost. Every call site passes a compile-time-constant mode and force_model is #[inline], so the branches constant-fold away — no extra call, no runtime dispatch. The RODAS4 Jacobian path skips the tide/GR/SRP/thrust terms (no partials), so it does not regress.
• Drag altitude gate uses norm_squared() against a precomputed squared limit instead of norm(), dropping a sqrt from every force eval.

2. Earth gravity: coefficient table capped at the evaluation degree

The evaluator dispatches at degree ≤ 40 (divisor tables are 44×44), so higher coefficients were never used — yet the default EGM96 model (file degree 360) stored a 361×361 DMatrix (~1 MB).

• Gravity::parse now caps the stored table at degree 44 (MAX_COEFF_DIM). Drops EGM96 to ~15 KB and shrinks the column stride 361 → 44, keeping the strided S-coefficient reads coeffs[(m-1, n)] resident in L1.
• Coefficients for n,m ≤ 43 are untouched → bit-identical results. Microbenchmark shows ~5–15% faster accel/accel_and_partials on EGM96 and removes the cache-miss timing variance in the hot loop.

Test plan

• cargo test full suite with SATKIT_DATA + test vectors — zero failures
• earthgravity unit tests 10/10
• Force-model results verified bit-identical (benchmark sink values matched across the change)
• cargo fmt --all clean
• Python suite not run (needs pip install -e . rebuild)

🤖 Generated with Claude Code