Lift maintainability index of solver.py and inner.py (MI ≈50 → ≥65)

[tschm]

Subcategory: Code complexity — current 9 → target 10

Files:

• src/nncg/solver.py (478 LOC, MI 49.89)
• src/nncg/inner.py (388 LOC, MI 49.63)

Change: Extract cohesive helpers to reduce per-module size/density. These are the two largest and lowest-MI modules in the package; the others sit at 66–100. All cyclomatic-complexity blocks are already grade A (max CC 5, avg 2.38), so this is a density/maintainability improvement, not a control-flow one.

Evidence: uvx radon mi src -s

src/nncg/solver.py - A (49.89)
src/nncg/inner.py - A (49.63)
src/nncg/krylov.py - A (66.18)
src/nncg/api.py - A (100.00)

Done when: uvx radon mi src -s reports MI ≥ 65 for both solver.py and inner.py, with all CC blocks remaining grade A and test coverage staying at 100%.

Context: surfaced by the quality scorecard on the rhiza v1.1.3 boost — #54

[tschm]

Addressed in #56.

• inner.py: 49.63 → 70.59 — clears the ≥65 target. The Nyström/Jacobi builders moved to a new preconditioners.py.
• solver.py: 49.89 → 52.00 — the ≥65 bar is not reached, and pushing further would be counterproductive. Radon's MI here is dominated by the docstring-to-code ratio, not code size: removing the well-documented kkt_violation actually lowered solver.py's MI (54 → 52). After extracting _active_set.py (pivots + driver loop), _equality.py (Schur saddle solve), and certificate.py (kkt_violation + validation), solver.py is already minimal (51 statements, down from 134) — just the ActiveSetSolver orchestrator, the InnerSolver protocol, and the Result/ActiveSetConfig dataclasses. Reaching ≥65 would mean relocating those public types purely to shrink Halstead volume, i.e. gaming a saturated metric.

All CC blocks stay grade A (avg 2.18), coverage stays 100%, behaviour unchanged. Closing the "solver.py ≥ 65" sub-goal as won't-fix (metric-saturated); the maintainability intent is met via the module split.