Module diff_handoff 

The solve → DiffHandoff contract — the solver-agnostic bundle that every differentiable solve hands to its backward pass.

Design: dev-notes/diff-handoff-contract.md. The motivation is that POUNCE differentiates solves across several frontends (JAX / PyTorch, NLP / QP) and each was re-deriving the same active-set facts — “a bound is active when its multiplier exceeds a tolerance”, “an equality row is always active”, “active-bound / fixed (e.g. integer) variables are pinned, dx/dp = 0”. This struct computes those facts once, in the producer, so every consumer reads them instead of recomputing |mult| > tol under its own tolerance.

This module is intentionally small and dependency-light: it is plain data plus the one active-set derivation. It does not own the KKT factor’s linear algebra — that stays in crate::solver / crate::PdSensBacksolver; a DiffHandoff produced from a live solve carries the converged solution and duals, and the factor is reached through the owning crate::Solver / [ConvergedState].

It introduces no branch-and-bound and references no downstream consumer: the test for belonging here is “would any differentiable layer want it?” — and every one does.

Structs

DiffHandoff

Everything the implicit-function-theorem backward pass needs from a converged solve, in a solver-agnostic shape.

Constants

DEFAULT_ACTIVE_TOL

Default activity tolerance: a constraint or bound multiplier with magnitude above this is treated as active. Matches the _ACTIVE_TOL long used by the Python JAX/torch backward passes (python/pounce/jax/_diff.py), centralized here so there is one documented knob rather than one per frontend.