feat(gds-psuu): parameter space constraints

[rororowyourboat]

Summary

Add declarative constraints to ParameterSpace so users can express feasibility regions beyond independent per-dimension bounds. Constraints are structural metadata — validated at construction time, inspectable by optimizers, and consistent with how every other GDS model handles invariants.

Motivation

Real parameter spaces often have cross-parameter dependencies: budgets that must sum to a limit, ratios that must stay positive, physical laws that couple variables. Currently, the only option is post-hoc filtering, which wastes evaluations and breaks optimizer assumptions.

Design

Follows the Pydantic @model_validator pattern used throughout GDS:

class Constraint(BaseModel, ABC):
    """Base class for parameter space constraints."""
    @abstractmethod
    def is_feasible(self, point: ParamPoint) -> bool: ...

class LinearConstraint(Constraint):
    """a₁x₁ + a₂x₂ + ... ≤ b"""
    coefficients: dict[str, float]
    bound: float

class FunctionalConstraint(Constraint):
    """Arbitrary feasibility predicate."""
    fn: Callable[[ParamPoint], bool]

ParameterSpace integration

ParameterSpace(
    params={
        "budget_a": Continuous(min_val=0, max_val=100),
        "budget_b": Continuous(min_val=0, max_val=100),
    },
    constraints=[
        LinearConstraint(coefficients={"budget_a": 1, "budget_b": 1}, bound=100),
    ],
)

Optimizer behavior

• GridSearchOptimizer: filter infeasible points from grid
• RandomSearchOptimizer: rejection sampling (with max retry limit)
• BayesianOptimizer: constrained acquisition function or rejection sampling
• ParameterSpace.is_feasible(point) — public method all optimizers can call

Acceptance criteria

• Constraint ABC with is_feasible()
• LinearConstraint and FunctionalConstraint implementations
• ParameterSpace.constraints field (default empty)
• ParameterSpace.is_feasible(point) checks all constraints
• Grid and random optimizers respect constraints
• Validation: constraint param names must be subset of space param names
• Tests with >90% coverage on new code