RFC: Decompose the four grade-D complexity hotspots

Problem

Radon rates exactly four functions at cyclomatic-complexity grade D — everything else in the package is C or better, so the complexity is concentrated and tractable:

DiscreteDP.__init__ alone dispatches dense/sparse × full/state-action-pairs input forms inline; _compute_fixed_point_ig interleaves algorithm and bookkeeping. High-CC plus hand-verified numerics is where regressions hide.

Proposed change

One PR per function; behavior-preserving extraction only:

• DiscreteDP.__init__ → extract _parse_full_form(...) / _parse_sa_form(...) validators returning a normalized bundle.
• _compute_fixed_point_ig → separate the iteration loop from error/logging bookkeeping.
• brent_max is Numba-jitted — some branching is inherent to Brent; treat as best-effort.
• Guard with before/after output equality on existing tests (coverage in these regions is strong).

Related context: #480 (repeated logic in several modules).

Acceptance criteria

• Radon grade ≤ C for the three non-jitted functions; no public API change
• Suite green with unchanged numerical outputs

From the July 2026 technical-debt audit (AI-assisted; claims verified against 28d4b3b on 2026-07-25).

Start by reading the four named entry points: _compute_fp.py:159, markov/ddp.py:299, optimize/scalar_maximization.py:5, and game_theory/howson_lcp.py:36, then run the existing suite and Radon. Work on one function per PR, preserving public APIs and numerical outputs; done means the three non-jitted functions reach Radon grade C or better and the suite remains green.