Add central-force potential family with conservation certificates

[xJoepec]

Summary

Generalizes the potential layer from the two hardcoded binding laws
(Harmonic, Coulomb) into an extensible central-force family, without
touching the integrator, the Landau–Lifshitz radiation-reaction reduction, or
the certificate discipline. New laws are a few lines each and inherit an exact
analytic force Jacobian.

The design goal was to extend the library in its own idiom: analytic forces
with analytic Jacobians (as the RR term requires), reduction oracles against the
existing trusted code, and re-checkable PASS/FAIL certificates on the physics.

What's added

Potentials (src/blueberry_circus/potentials.py)

• CentralPotential base: a subclass supplies only three radial functions
  U(s), U'(s), U''(s) of the softened radius s = √(r² + softening²); the
  base derives the Cartesian force and an analytic force_jacobian via
  F = g(s) x, g = −U'/s, J = g I + (g'/s) x xᵀ.
• PowerLaw, Yukawa, Morse, AnharmonicOscillator (isotropic Duffing),
  LennardJones. PowerLaw reduces exactly to Harmonic (p=2) and
  Coulomb (p=−1) — used as reduction oracles in the tests.

Declarative ops (src/blueberry_circus/program.py)

• Matching Operations for each law (bc.PowerLaw, bc.Yukawa, bc.Morse,
  bc.Anharmonic, bc.LennardJones).
• Program.compile() now collects potentials by a _PotentialOp marker instead
  of a hardcoded (Harmonic, Coulomb) tuple, so future potentials are picked up
  automatically (the exhaustiveness pass keys on the same marker).

Conservation certificates (src/blueberry_circus/conservation.py)

• energy_conservation_certificate and
  angular_momentum_conservation_certificate encode the fractional peak-to-peak
  drift of the invariant against the canonical residual_le_tol rule. An
  under-resolved integration or a tampered trajectory re-derives FAIL. For a
  central force, |L| conservation is the signature of centrality, and energy
  conservation cross-checks that a new potential's analytic force is consistent
  with its potential energy — purely from the integrated dynamics.

Docs & example: docs/central_forces.md, examples/central_forces.py.

Testing

tests/test_central_potentials.py (33 cases) pins correctness three independent
ways:

1. Finite differences — analytic force == −∇U and force_jacobian == dforce/dx;
   Jacobian symmetry (curl-free).
2. Reduction oracles — PowerLaw reproduces Harmonic and Coulomb to
   floating-point closeness.
3. Conservation — closed orbits pass the energy and |L| certificates (all
   six families to ≲10⁻¹¹ at dt = 5·10⁻³), and a real drift FAILs an
   impossibly tight tolerance.

Plus radiation-reaction stability for the new families and declarative-program
integration through the Engine.

$ pytest -q
167 passed, 2 xfailed

No changes to existing behavior: Harmonic/Coulomb and all prior certified
results are untouched (the PowerLaw reduction tests prove the generalization is
consistent with the trusted core).

Motivation

This is the enabling substrate for a general dynamics corpus in which
BlueberryCircus is the data-generating engine: continuous force-law families
(via PowerLaw's exponent and the qualitatively distinct Yukawa/Morse/Duffing/LJ
wells) plus certified invariants give clean, re-checkable ground truth for
inverse-dynamics and law-inference tasks.