fix(mental-models): full refresh pending delta baselines

[xmh1011]

Summary

• Treat the mental-model creation placeholder Generating content... as pending content, not as a valid delta baseline.
• Make the first delta-mode refresh fall back to full recall/synthesis so pre-existing scoped facts are considered.
• Add a regression test asserting the placeholder path does not pass created_after or call the delta-merge LLM.

Closes #1679.

Tests

• HINDSIGHT_API_LLM_PROVIDER=none uv run pytest tests/test_mental_model_delta.py::TestDeltaRefreshPlumbing -q
• uv run ruff check hindsight_api/engine/memory_engine.py tests/test_mental_model_delta.py
• uv run ty check hindsight_api/engine/memory_engine.py
• ./scripts/hooks/lint.sh

[yugandhar-maram]

Thanks for the quick fix!