Add DiscreteBatchConstraint

[Scienfitz]

Closes #583
Fixes #567

Functionality:
Adds a DiscreteBatchConstraint which works via subspace generation just like the continuous cardinality constraint, just in the discrete searchspace part. While it will quickly lead to many partitions, it is also possible to use multiple DiscreteBatchConstraints and also combine them with the cardinality constraints because there is now a more unified interface for constraints that work via subspace generation.

Notes:

• Works identical to cardinality, except it operates on discrete space and generates discrete partitions
• The naming (originally taken form the cardinality work) using subspace is not optimal, because there is also the concept of DiscreteSubspace / ContinuousSubspace. I changed the naming that references to creating subspaces for taking care of the constraint into partition. So we would then have things like partition_masks, n_theoretical_partitions etc
• Discrete partitions are identified via masks, while continuous partitions remain identified by parameter names
• The hyrbid case was now implemented because it is now possible to have discrete and continuous partition-generating constraints at the same time. There is nothing conceptual that speaks against this although the number of subspaces possible explodes due to Cartesian product of individual subspaces. The limit for this is n_max_partitions which controls the max number of subspaces considered in any case
• The methods that create the objects that identify the partitions are coded to do this lazily and return iterators. They include arguments to do that indefinitely (replace=True) or shuffle the order (shuffle=True). The latter is used when subsamapling the spaces because with shuffle=True you can just instantiate the fist n objects in the iterator to achieve random subsampling. This has some complications, especially for the overall searchspace class which needs to have a parameter to stop the process when replicated samples are drawn too often (likely no other feasible subspace combinations).
• botorch.py in recommenders has become very large as a result of this development. So I split it into submodules
• _FixedNumericalContinuousParameter had a bugged property which was named is_numeric but it should be is_numerical
• Since hybrid is now implemented, the ContinuousCardinalityConstraint can now also be applied in hybrid spaces, which fixes ContinuousCardinalityConstraint not considered in hybrid spaces #567

Discuss:

1. n_theoretical_subspaces just computes the upper limit of subspaces in discrete and hybrid cases. Because subspaces that do not have enough candidates are internally skipped the actual number of n_feasible_subspaces might be smaller. However that cannot be easily computed and would require instantiating all masks corresponding to the subspaces. The dispatcher logic is currently based on n_theoretical_subspaces but to be 100% correct it should be n_feasible_subspaces - this is not implemented due to the mentioned difficulty. I think using n_theoretical_subspaces works as a proxy and will not be very different from n_feasible_subspaces for nearly all practical problems

[Copilot]

Pull request overview

Adds a new batch-level discrete constraint (DiscreteBatchConstraint) that enforces a fixed discrete parameter value across all points in a recommended batch by optimizing over discrete (and hybrid) subspaces, aligning with the existing subspace-based infrastructure used for continuous cardinality constraints.

Changes:

• Introduces DiscreteBatchConstraint (non-filtering, batch-level) plus validation preventing duplicate batch constraints per parameter.
• Extends discrete/hybrid subspace infrastructure (n_theoretical_subspaces, mask/config iterators, sampling helpers) and wires support into BotorchRecommender and RandomRecommender with compatibility gating for other recommenders.
• Adds documentation and test coverage for discrete + hybrid subspace-generating constraints, replacing prior hybrid-only cardinality tests.

Reviewed changes

Copilot reviewed 14 out of 14 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
baybe/constraints/discrete.py	Adds DiscreteBatchConstraint and its subspace mask generation.
baybe/constraints/validation.py	Adds validation to prevent multiple batch constraints on the same parameter.
baybe/constraints/__init__.py	Exposes DiscreteBatchConstraint publicly.
baybe/searchspace/discrete.py	Adds discrete subspace-generating constraint plumbing and mask enumeration/sampling helpers.
baybe/searchspace/continuous.py	Renames/aligns to n_theoretical_subspaces and adds shuffled / with-replacement subspace configuration iteration.
baybe/searchspace/core.py	Adds combined discrete+continuous theoretical subspace counting and combined sampling utilities for hybrid spaces.
baybe/recommenders/pure/base.py	Adds supports_discrete_subspace_constraints gate and raises IncompatibilityError when unsupported.
baybe/recommenders/pure/bayesian/botorch.py	Implements discrete + hybrid optimization over subspaces when batch constraints are present.
baybe/recommenders/pure/nonpredictive/sampling.py	Updates RandomRecommender to respect discrete subspace-generating constraints.
docs/userguide/constraints.md	Documents DiscreteBatchConstraint and recommender compatibility.
tests/constraints/test_batch_constraint.py	Adds focused tests for the new constraint (behavior, validation, incompatibility, subspace counting/masks).
tests/constraints/test_subspace_constraints_hybrid.py	Adds parametrized hybrid tests covering discrete batch + (discrete/continuous) cardinality constraints.
tests/constraints/test_cardinality_constraint_hybrid.py	Removes older hybrid cardinality-only test file (superseded).
CHANGELOG.md	Records new user-facing feature entry.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.