Add an energy-audited SED hydrogen hypothesis tournament

[xJoepec]

What

• compute the complete Nieuwenhuizen point-charge drift surface D(E,L) before applying any modified model, using a perihelion-adapted improper quadrature and analytic short-history cancellation
• add OrbitState, EnergyLedger, TournamentConfig, and HypothesisResult, including re-derived ledger closure, stored seeds, coupling scales, resolutions, and 95% confidence intervals
• implement the four preregistered arms: canonical SetterfieldDrive, reciprocal FiniteShellResponse, InverseSquareControl, and a conservative Rodriguez-inspired MultipoleStorage surrogate
• enforce the closed classification vocabulary CHANNEL_SUPPRESSED, ACTIVE_CONTROL, NO_EFFECT, DESTABILIZED, and NULL
• add an explicit, chunkable blueberry-tournament research command with deterministic JSON summaries and optional NPZ arrays; ordinary CI runs only a reduced schema and ledger smoke test
• release the package and Rust crate as 0.3.0

Scientific scope and source audit

The finite-mode layer is a stochastic quadrature of the perturbative response kernel, not a physical-timescale nonlinear hydrogen trajectory. Rectification remains conditional energy-space drift, not equilibrium vacuum-energy extraction or a stationary ground-state claim.

Direct quadrature of the kernel printed in Nieuwenhuizen 2016, Eqs. 56 through 64, gives H_max approximately 7.327 at mu approximately 0.590, hence the defining d_c=-H_max^2 approximately -53.69. The paper prose value -35.8 squares only its quoted endpoint H(0)=5.99, despite stating an all-mu criterion. Reports retain both numbers and do not hard-code either as empirical physics. The implementation independently recovers G(1)=3 L_c/2, tying that source kernel back to the certified point-charge radial limit.

Setterfield and the Rodriguez-inspired internal mode remain explicitly speculative hypotheses. Any negative drift with nonzero Setterfield parameter work is classified as active control. The multipole oscillator is a Hamiltonian surrogate, not a validated proton model.

Verification

• core: 134 passed, 2 strict xfailed, 0 skipped; 136 selected from 147 collected
• Rust backend: 4 passed after building blueberry_rs 0.3.0
• JAX backend: 4 passed
• all five examples passed, including the reduced four-arm schema and ledger smoke
• isolated wheel install from /tmp: blueberry-circus==0.3.0, neutral-directory import, public tournament dataclasses, and console-script manifest all passed
• ruff check, compileall, and git diff --check passed

No certificate rule was weakened, no mandatory check was skipped, and no automation or co-author trailer is present.