Forward-mode AD via the shared scalar rule table

[dpsanders]

Summary

Stacked on #83. Adds forward-mode AD, reusing the unary_rule / binary_rule dispatch table that PR #83 introduced for reverse-mode. One rule table, two AD passes — no second source of truth to drift.

API

• dotted(v) — tangent variable for v, written with a combining overdot (parallel to the internal bar(v) for reverse mode).
• tangent_code(ex_or_ssa, vars) — appends one tangent assignment per forward assignment, surfacing input_tangents and output_tangent in SSAFunction.variables.
• tangent(ex, vars) — exported. Compiles a callable

  (__inputs, __tangents) -> (value, directional_derivative)

Behaviour

@variables x y
t = tangent(x^2 + y^2, [x, y])
t((3.0, 4.0), (1.0, 0.0))   # (25.0, 6.0) — ∂f/∂x = 2x
t((3.0, 4.0), (0.0, 1.0))   # (25.0, 8.0) — ∂f/∂y = 2y
t((3.0, 4.0), (0.7, -0.4))  # full Gateaux derivative along (0.7, -0.4)

Verified against reverse-mode gradient in tests.

Note on RGF naming

The compiled function's first argument is __inputs, not __args. The latter collides with RuntimeGeneratedFunctions' own internal binding and triggers a spurious BoundsError at call time. Documented inline where the Expr is built.

Test plan

• 68/68 tests pass (59 inherited from Derivative hardening: own the scalar rule table #83, +9 new for tangent).
• Directional derivative matches reverse-mode gradient component-by-component.
• Full Gateaux derivative along arbitrary direction matches ⟨∇f, v⟩.

Follow-ups

Hessian via forward-over-reverse (applying tangent_code to the SSA produced by gradient_code) will land in a separate PR on top of this one.

🤖 Generated with Claude Code