Four missing/incomplete pieces of the TensorProgram learning pipeline, all natural follow-ons to the gradient computation work in the previous PR.

Changes

TensorEquation.h / .cc

• Config accessors — _max_iterations and _convergence_threshold were private with no public API; adds set_max_iterations(), max_iterations(), set_convergence_threshold(), convergence_threshold()
• Gradient clipping — adds set_grad_clip(double) / grad_clip(); when non-zero, each gradient vector is L2-norm-clamped in update_parameters() before the SGD step; extracted into a clip_gradient_vector() static helper to avoid duplication
• Loss history — train() now clears and appends to _loss_history each epoch (pre-update, standard ML convention); exposed via loss_history() -> const std::vector<double>&
• HYBRID mode — execute() previously treated HYBRID identically to CONTINUOUS; now applies sigmoid (when no explicit nonlinearity is set) then threshold, yielding binary outputs with gradient flow via the existing straight-through estimator in backward()

TensorProgram prog("learn");
prog.set_learning_rate(0.01);
prog.set_grad_clip(1.0);          // prevent exploding gradients
prog.set_max_iterations(50);      // limit fixpoint iterations
prog.set_convergence_threshold(1e-4);

prog.train({}, targets, 100);

// Inspect per-epoch loss
for (double loss : prog.loss_history())
    std::cout << loss << "\n";

Tests

10 new tests: max_iterations / convergence_threshold accessors and their effect on forward_to_fixpoint(); grad_clip round-trip and that clipping actually bounds weight deltas to ≤ lr × clip; loss_history length equals epoch count, values non-negative, and loss decreases over training; HYBRID mode produces binary {0,1} output, routes through sigmoid→threshold when nonlinearity is NONE, and skips the extra sigmoid when an explicit nonlinearity is set.

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.