v0.2.1

0.2.1 (2026-07-22)

Features

• Add ConvNN, NNConvParams, and Nets.CONV (#89).
• Add FeedFwdMoENN, NNMoEParams, and Nets.FEED_FWD_MOE (#88).
• Add full-batch graph node-classification loading without pyg-lib.
• Add per-layer activation and dropout configuration (#85).
• Add pluggable eval_step_fn validation hooks (#86).

Bug Fixes

• Persist callback-driven network mutations in the final checkpoint (#87).
• Reconstruct converted QAT checkpoints from saved transform metadata.
• Harden training lifecycle, parameter validation, and pandas 2 compatibility.
• Refresh the dependency lock to remove the vulnerable GitPython resolution.
• Complete specialized network exports and reject invalid ConvNN/MoE configurations.

Documentation

• Add and update runnable examples for the new model and training surfaces.
• Link the rendered documentation site and provide a themed architecture view.
• Synchronize architecture and publishing documentation.