Adds failing-first Behave regression scenarios proving issue #122:
Executor.__init__ never calls AgentFactory.validate_configuration(),
so a malformed agents.<name> entry (missing the required "type" key,
or a non-mapping value) is accepted silently instead of being
rejected per docs/index.md §11.1, which requires agent validation to
be enforced before any agent is instantiated or message processed.

Each scenario builds a config for a single-node graph route whose
only agent entry is malformed, then constructs and executes the
Executor in one step, capturing any exception raised during either
construction or execution. This mirrors the established pattern in
pure_graph_edge_target_validation_steps.py and keeps the capture
robust regardless of whether the eventual fix for #122 validates
inside Executor.__init__ (as #122's acceptance criteria require) or
elsewhere in the dispatch path -- avoiding a scenario that would
break outright (rather than correctly flip to passing) once the fix
lands, since @tdd_expected_fail only inverts AssertionError and a
construction-time ConfigurationError would not be one.

Confirmed locally that both assertions fail against current master
when @tdd_expected_fail is removed: the malformed entries instead
default to an empty type: llm agent and fail several layers
downstream with an unrelated "missing credentials for provider" or
attribute error, exactly matching the bug's reported symptom. Tagged
@tdd_issue @tdd_issue_122 @tdd_expected_fail so CI inverts the
expected failure until the fix lands on a bugfix/ branch.

Refs: #122
ISSUES CLOSED: #124

@@ -97,0 +134,4 @@
+            }
+        },
+        "agents": {
+            agent_name: "not-a-mapping",