fix(tests): make mock retain extraction deterministic

[JulienJBO]

Summary

• generate deterministic sentence-level fact extraction output for the mock LLM when retain uses scope=retain_extract_facts
• default the shared memory test fixture to the mock provider when no local LLM env is configured
• add coverage for the new mock retain behavior

Why

Running hindsight-api-slim tests locally without a populated .env currently falls back to a real provider (groq) and fails early, or, when forced to mock, retain-heavy integration tests extract zero facts and produce large misleading failure clusters. This keeps the generic mock behavior intact while making local retain/recall integration tests deterministic and useful.

Validation

• uv run pytest tests/test_per_operation_llm_config.py::TestMockLLMProvider::test_mock_provider_generates_facts_for_retain_extract_scope -q
• uv run pytest tests/test_retain_append_mode.py::test_append_mode_concatenates_content -q
• uv run pytest tests/test_tags_visibility.py -q
• uv run ruff check hindsight_api/engine/providers/mock_llm.py tests/conftest.py tests/test_per_operation_llm_config.py

Follow-up

A broader local no-env test pass still has a second cluster around richer mock behaviors (consolidation, causal_relations, entity_labels, and zero-fact edge cases). I kept this PR scoped to the first large retain/recall failure wave.