Encode backend-neutral semantic flat inputs

Use SemanticFlatRelationEncoderEngine when planning data already has an owned semantic representation and should not be tied to a Mimir domain, problem, or state object. The engine implements the same flat relation, target, history, predicate-virtual, LGAN, and batch contract as FlatRelationEncoderEngine, but its input contains only strings and integer indices.

The engine is initialized once from predicate and action schema:

import mifrost

category = mifrost.SemanticPredicateCategory
predicates = [
    mifrost.SemanticPredicateSpec(category.fluent, "at", 2),
]
actions = [
    mifrost.SemanticActionSpec("move", 3),
]
engine = mifrost.SemanticFlatRelationEncoderEngine(predicates, actions)

Each graph owns its object table. Predicate indices refer to the engine's predicate vector, action indices refer to its action vector, and object indices refer to that graph's object vector:

value = mifrost.SemanticFlatRelationInput()
value.objects = ["robot", "room-a", "room-b"]
value.state_facts = [mifrost.SemanticAtom(0, [0, 1])]
value.goals = [
    mifrost.SemanticLiteral(mifrost.SemanticAtom(0, [0, 2]), True),
]
value.actions = [mifrost.SemanticGroundAction(0, [0, 1, 2])]

encoding = engine.encode(value)
tensors = encoding.as_pyg()

These indices are compact call-local references, not persistent semantic IDs. For process, repository, replay, or checkpoint boundaries, serialize names and reconstruct the local indices deterministically. Never serialize planning repository indices or assume independently constructed repositories share identity.

encode_batch accepts any number of compatible owned inputs, including an empty list. It uses the same node offsets, LGAN relation-instance offsets, target-local candidate IDs, and optional relation-major packing as the Mimir engine:

config = mifrost.FlatRelationEncoderConfig(
    pack_relation_args_relation_major=True,
)
engine = mifrost.SemanticFlatRelationEncoderEngine(predicates, actions, config)
batch = engine.encode_batch([value, value])

Input validation rejects duplicate schema/object names, invalid arities, out-of-scope indices, non-negative history deltas, unsupported target sources, and subgoal layers beyond max_goal_level before emitting a graph.