docs: index.md and architecture_overview.md claim Jacobians use JAX autodiff but production uses OpenVAF analytical Jacobians

[github-actions]

Problem

Two docs describe Jacobian computation as using JAX autodiff (jacfwd/jvp), but the production hot path uses analytical Jacobians generated by OpenVAF's symbolic differentiation.

Affected locations

docs/index.md line 5:

• Automatic differentiation for computing device Jacobians without explicit derivatives

docs/architecture_overview.md line 8:

2. Automatic Differentiation: Jacobians computed via JAX autodiff, no explicit derivatives

docs/architecture_overview.md lines 54–57 (diagram):

│     │   Autodiff         │
│     │ (Jacobians via     │
│     │  jacfwd/jvp)       │

Why it's wrong

vajax/analysis/homotopy.py line 16 explicitly states:

"uses analytic jacobians from OpenVAF. This avoids jax.jacfwd() autodiff"

The production transient solver path (solver_factories.py) receives Jacobians directly from OpenVAF's MIR-compiled functions. JAX's jacfwd/jvp is only used in analysis/solver.py (the DC utility solver), not in the transient hot path.

docs/gpu_solver_jacobian.md documents the full history of this migration from autodiff to analytical Jacobians.

Fix

Update the design principle description in both files to say something like:

Analytical Jacobians: Jacobians computed symbolically by OpenVAF's MIR compiler — no runtime autodiff overhead

The diagram in architecture_overview.md should replace the "Autodiff (Jacobians via jacfwd/jvp)" box with "Analytical Jacobians (via OpenVAF MIR)".