Problem

User preferences, rules, corrections, and constraints can be stated during assistant interactions, but they are durable facts about the user rather than assistant experiences. The current fact extraction guidance made that boundary easy to blur, which can send durable user facts through the experience path.

Fixes #2437.

Fix

• Clarify the retain fact_type prompt so user preferences/rules/corrections/constraints remain world even when mentioned in an assistant interaction.
• Narrow assistant to actions or experiences the assistant/agent actually performed.
• Align the Pydantic schema field descriptions and verbatim/verbose prompt guidance with the same boundary.
• Add a prompt/schema regression test that locks this wording down without calling a live LLM.

Test

• uv run pytest tests/test_fact_extraction_fact_type_prompt.py -q
• uv run ruff check hindsight_api/engine/retain/fact_extraction.py tests/test_fact_extraction_fact_type_prompt.py
• uv run ruff format --check hindsight_api/engine/retain/fact_extraction.py tests/test_fact_extraction_fact_type_prompt.py
• git diff --check

Risk

Low-to-medium. This is prompt/schema guidance only, not persistence or mapping logic. It should reduce misclassification of user facts, but prompt wording can affect extraction behavior, so the change is intentionally narrow.