Expose LLM sampling temperature as an admin-configurable setting #70

Context

docs/CONFIG_AUDIT_2026-09-04.md found that backend/app/core/llm_client.py:94 hardcodes temperature=0.7 on the ChatOpenAI construction, with no env var or DB override at all — conspicuously the one LLM parameter left out of the otherwise-complete resolve_llm_config resolver, which already covers provider/url/model/timeout via system_settings_service.py.

Scope

1. Add nullable SystemSettings column llm_temperature (float), env-default fallback in backend/app/core/config.py, resolved via the existing resolve_llm_config (extend it — this is the same resolver, not a new one).
2. Wire into llm_client.py's ChatOpenAI construction, replacing the literal 0.7.
3. Add to the existing Settings UI LLM section (no new section needed) with a tooltip explaining the tradeoff (lower = more deterministic/factual, higher = more varied).
4. Audit-log on change, folded into the existing LLM config audit event if one exists, or "llm_config.temperature_update" if scalar fields are audited individually.

Definition of Done

• llm_temperature nullable column + Alembic migration.
• resolve_llm_config extended, unit-tested.
• Wired into llm_client.py.
• Settings UI field + tooltip added to the existing LLM section.
• Audit logging on change.
• RBAC call-out per AGENTS.md TDD Discipline step 0: admin feature.
• Standard merge gate from AGENTS.md's Definition of Done applies in full.