on_mutation_outcome forwards program.get_metadata('mutation_model') directly to self._bandit.update_reward, which indexes self.arms[arm_name]. The normal flow keeps producer and router in step, but the metadata can also arrive from a Program loaded out of a Redis snapshot written by an older router, from a custom mutation operator, or from a hand-built test fixture. In those cases update_reward raises KeyError and aborts the agent callback.

Minimal reproducer against current main:

from unittest.mock import MagicMock
from gigaevo.llm.bandit import BanditModelRouter, MutationOutcome
from gigaevo.programs.program import Program

def _mock(name):
    m = MagicMock(); m.model_name = name
    m.with_structured_output = MagicMock(return_value=MagicMock())
    return m

router = BanditModelRouter([_mock('llama'), _mock('qwen')], [0.5, 0.5], fitness_key='score')
child = Program(code='x=1')
child.set_metadata('mutation_model', 'gpt-4-not-in-router')
child.metrics['score'] = 0.8
parent = Program(code='x=0'); parent.metrics['score'] = 0.5
router.on_mutation_outcome(child, [parent], outcome=MutationOutcome.ACCEPTED)
# KeyError: 'gpt-4-not-in-router'

The change adds a single membership check against self._bandit.arms at the top of on_mutation_outcome and skips with one debug note when the name is unknown. Three new tests cover the unknown-arm path on the ACCEPTED branch, on the REJECTED_ACCEPTOR branch, and combined with missing child fitness.