QuadprogProjector does not validate norm_eps and reg_eps at construction #810

The norm_eps and reg_eps setters of QuadprogProjector raise a ValueError on negative values, but __init__ assigns self._norm_eps and self._reg_eps directly, so the check is skipped at construction:

>>> from torchjd.linalg import QuadprogProjector
>>> p = QuadprogProjector(norm_eps=-1.0, reg_eps=-5.0)  # accepted
>>> p.norm_eps = -1.0
ValueError: norm_eps must be non-negative, but got -1.0.

A negative reg_eps is subtracted from the diagonal of the Gramian, which can make the quadratic program non-convex. The user then gets a "Failed to solve the quadratic programming problem." error much later, instead of a clear error when creating the projector. The aggregators (e.g. CAGradWeighting) already go through their setters in __init__.

Suggested fix

super().__init__()
self.norm_eps = norm_eps
self.reg_eps = reg_eps

Checklist

• Use the setters in QuadprogProjector.__init__.
• Add a test in tests/unit/linalg/test_dual_cone.py checking that QuadprogProjector(norm_eps=-1.0) and QuadprogProjector(reg_eps=-1.0) raise a ValueError.

See CONTRIBUTING.md to set up the development environment.