fix(mental-models): align stale checks with refresh tag scope

[r266-tech]

Summary

• make mental-model staleness use the same resolved tag filtering as refresh
• preserve tagged-model all_strict defaults and explicit tags_match behavior
• support compound trigger.tag_groups when deciding whether a scheduled refresh is stale
• keep fact_types as an additional scope constraint

Why

The scheduled-refresh gate and the refresh itself currently resolve tag scope differently. Refresh uses _resolve_refresh_tag_filtering(), including tag_groups and the tagged-model all_strict default, while compute_mental_model_is_stale() ignores tag_groups and defaults tagged models to any.

That can mark a model stale because of an out-of-scope memory, spend an unnecessary refresh LLM call, and then refresh against a different (or empty) corpus. This was surfaced while tracing #2800.

Testing

• uv run pytest tests/test_mental_models.py::TestMentalModelStaleness -q -n 0 (12 passed)
• uv run ruff format hindsight_api/engine/memory_engine.py tests/test_mental_models.py
• uv run ruff check hindsight_api/engine/memory_engine.py

The regression coverage includes default all_strict, explicit any, compound tag groups, untagged/global behavior, and flat/group tag filters composed with fact_types, including multi-group near misses.

Related to #2800.

[koriyoshi2041]

Verified this at d0fbf450 on macOS / Python 3.11 with the embedded Postgres and local embedding/reranker fixtures:

• uv run pytest tests/test_mental_models.py::TestMentalModelStaleness -q -n 0 — 12 passed
• uv run ruff check hindsight_api/engine/memory_engine.py — passed
• uv run ruff format --check hindsight_api/engine/memory_engine.py tests/test_mental_models.py — passed

The added cases cover the important scope boundaries: default all_strict, explicit any, compound tag_groups, and composition with fact_types. The stale gate now uses the same resolved filtering helpers as refresh, so an out-of-scope memory no longer schedules a refresh against a different corpus.

One scope note for #2800: this fixes stale-check/refresh disagreement, but it intentionally does not change the refresh recall semantics that produced the reporter's zero-result case; the documented tags_match: any or narrower model tags remain the relevant configuration path there.