RL Numeric Diagnostics

This umbrella collects optional numeric hardening tools for runtime RL code. These modules do not replace the ordinary scalar-polymorphic RL equations; they add checked execution paths for cases where we want explicit floating-point semantics or interval-style diagnostics.

Current contents:

• NN.Runtime.RL.Numerics.Float32.Types: checked Float → IEEE32Exec casting.
• NN.Runtime.RL.Numerics.Float32.Returns: explicit binary32 discounted backups and returns.
• NN.Runtime.RL.Numerics.Float32.Advantage: TD residuals, GAE, and normalization with finite-intermediate checks.
• NN.Runtime.RL.Numerics.Float32.PPO: checked PPO scalar objective helpers.
• NN.Runtime.RL.Numerics.Float32.Intervals: outward-rounded interval enclosures.

References:

• IEEE 754-2019, floating-point arithmetic.
• IEEE 1788-2015, interval arithmetic.
• Goldberg, "What Every Computer Scientist Should Know About Floating-Point Arithmetic", 1991.