feat: H_φγ invariant — energy ratio convergence analysis at γ≈1

[Copilot]

Tests whether attention_ratio(t) = energy_core(t) / energy_periphery(t) converges to φ≈1.618 or φ⁻¹≈0.618 when γ(t)→1.0 across neosynaptex substrates, using three independent ratio definitions and three null models.

Core analysis (analysis/phi_gamma_invariant.py)

• R1 — topology-energy: core/periphery partition by weighted-degree centrality (top 38.2%)
• R2 — spectral-core: dominant PSD modes covering top 38.2% spectral energy
• R3 — predictive-allocation: features with top 38.2% AR(1) prediction weight
• Gamma windows via Welch PSD + Theil-Sen slope in log-log space
• Bootstrap CI on median ratio, permutation p-value against null distribution
• Verdict logic: support | reject | insufficient with explicit gates (n≥30, CI width, p<0.05)

Null models (analysis/phi_gamma_nulls.py)

• Temporal shuffle — destroys temporal order, preserves amplitudes
• Topology shuffle — random core/periphery reassignment, preserves group sizes
• Phase randomization — preserves power spectrum, destroys phase correlations

Report (analysis/phi_gamma_report.py)

• JSON schema with per-substrate entries (all 3 ratio methods) + cross-substrate aggregation
• Output: evidence/phi_gamma/phi_gamma_report.json

CLI

python -m analysis.phi_gamma_invariant \
  --config experiments/phi_gamma_config.yaml \
  --out evidence/phi_gamma/phi_gamma_report.json

Tests — 24 cases covering:

• φ constant precision, φ·φ⁻¹=1 identity
• R(t) finiteness, delta guard (periphery energy = 0)
• Unity window selection with IEEE 754 boundary behavior
• All three verdict states constructed explicitly
• Null model invariants (size preservation, power spectrum preservation)
• Bootstrap CI ordering and coverage
• Full pipeline end-to-end