Fix loop export reference inference and Jinja template resolution for loop contexts

[zhongkaifu]

Motivation

• Tests were failing because references to loop outputs sometimes omitted the exports segment and validation rejected inferred paths.
• Nested loop source checks incorrectly treated loop/loop_ctx references as non-array sources during model validation.
• Template evaluation did not resolve bare {{ ... }} reference paths against the binding context early, causing length/count and built-in function usages to return incorrect values.

Description

• Make BindingContext tolerant of both wrapped (result_of.<loop>.exports.<key>) and flattened (result_of.<loop>.<key>) loop result shapes by skipping the exports segment when the runtime result is flattened and by improving the inference path logic in resolve_binding/_get_from_context.
• Adjust BindingContext template handling so a bare Jinja template {{ ... }} that looks like a simple reference to loop/loop_ctx/result_of is resolved via ctx.get_value before generic Jinja evaluation, and fall back to the context value when a Jinja Undefined is produced.
• Skip loop source schema checks for paths that are loop or start with loop./loop_ctx. in Workflow.validate_loop_body_subgraphs to avoid false positives for loop-local references.

Testing

• Reproduced the original failures and ran the focused test set covering loop binding inference and template resolution.
• Executed pytest -q tests/test_reference_templates.py::test_reference_allows_builtin_function_invocation tests/test_reference_templates.py::test_eval_params_renders_interpolated_templates tests/test_loop_binding_inference.py::test_missing_exports_segment_is_inferred tests/test_loop_state_isolation.py::test_nested_loop_results_do_not_leak_across_outer_iterations tests/test_reference_templates.py::test_binding_context_resolves_length_field tests/test_reference_templates.py::test_binding_context_resolves_count_field and all tests passed.

---

Codex Task