Expose rate-limit bucket idle-eviction window as admin-configurable #73

Context

docs/CONFIG_AUDIT_2026-09-04.md found that backend/app/services/rate_limiter_service.py:58 hardcodes _STALE_AFTER_SECONDS = 3600 — governing when an idle in-process rate-limit bucket entry (keyed per user_id or IP) is evicted, to bound memory growth. This sits right next to the already-DB-backed rate-limit capacity/refill fields (rate_limit_chat_capacity, rate_limit_chat_refill_per_minute, etc., Phase 4f) but wasn't included in that config surface.

Why this matters

An operator running with a much lower refill rate than the eviction logic's assumed baseline (the code comment assumes roughly 1 token/min) might want this window raised so buckets aren't evicted before they'd naturally refill, or lowered to bound memory more aggressively on a high-cardinality IP-keyed endpoint (POST /api/auth/token, unauthenticated and attacker-influenced per Phase 4f's hardening notes).

Scope

1. Add nullable SystemSettings column rate_limit_bucket_stale_after_seconds, following the exact pattern of the existing rate-limit fields in rate_limit_config_service.py.
2. Wire into rate_limiter_service.py, replacing the module constant.
3. Add to the existing Settings UI "Rate Limiting" section (added in the Settings Screen Help Text work) with a tooltip.
4. Audit-log on change, folded into the existing "rate_limit.config_update" event.

Definition of Done

• rate_limit_bucket_stale_after_seconds nullable column + Alembic migration.
• resolve_rate_limit_config extended, unit-tested.
• Wired into rate_limiter_service.py's eviction logic.
• Settings UI field + tooltip added to the existing Rate Limiting section.
• Audit logging on change (existing rate_limit.config_update event).
• 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.