fix(s3): reject responses with incomplete tags

[Iams4kura]

Summary

• Fixes: BehaviorNarrator accepts plain text or a response missing either the thoughts or answer tag pair, so call_llm_formatted skips its corrective retry and malformed caption data reaches bBoN trajectory comparison.
• Root cause: split_thinking_response uses the last split segment when a tag is absent, which makes the full untagged response look like parsed thoughts and answer text; thoughts_answer_tag_check only checked that the parsed thoughts value was nonempty instead of verifying that all four tags exist in order.

Regression evidence

• Before: UV_CACHE_DIR=/tmp/agent-s-python311.20260821T160739Z/cache uv run --python 3.11 --no-project --with pillow -- python -m unittest discover -s tests -p 'test_formatters.py' -v exited 1

• After: UV_CACHE_DIR=/tmp/agent-s-python311.20260821T160739Z/cache uv run --python 3.11 --no-project --with pillow -- python -m unittest discover -s tests -p 'test_formatters.py' -v exited 0

Verification

• UV_CACHE_DIR=/tmp/agent-s-python311.20260821T160739Z/cache uv run --python 3.11 --no-project --with pillow --with numpy --with backoff --with openai --with anthropic -- python -m unittest discover -s tests -v
• UV_CACHE_DIR=/tmp/agent-s-python311.20260821T160739Z/cache uv run --python 3.11 --no-project --with black -- black --check gui_agents tests/test_formatters.py

Scope

• 2 files changed, +39 / -1 lines