fix: 종료된 deepseek-v4-flash:cloud → gemma4:31b-cloud (기억 재랭킹·Google 폴백)

Ollama Cloud가 2026-09-25에 deepseek-v4-flash를 종료해 재랭킹이 HTTP 410으로 실패하고
거리 임계값 폴백만 쓰이던 문제.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

src/gateway/memory/memory-rerank.ts

@@ -14,7 +14,7 @@ import { getConfig } from '../../config/config.js';
 // "flash"-tier model chosen specifically for latency, not quality — this call is a small
 // classification task (pick relevant indices from a short list), not a task worth spending
 // primary-model-grade reasoning time on. Revisit if this model is ever unpulled/renamed.
-const RERANK_MODEL = 'deepseek-v4-flash:cloud';
+const RERANK_MODEL = 'gemma4:31b-cloud';
 // Only fires on genuinely ambiguous retrieval (see the gating in personality-context.ts) —
 // most turns never reach this call at all, so a generous timeout here doesn't cost anything
 // on the common path. 8s was measured to occasionally clip real (successful, just slow)

src/providers/factory.ts

@@ -28,7 +28,7 @@ import { isGoogleNearLimit } from './google-usage';
 // reactive to a 429 (see google-usage.ts). Hardcoded rather than reusing
 // models.fallback: that field is a separate, more general "provider errored
 // mid-turn" mechanism (see handle-chat.ts) and may be tuned independently.
-const GOOGLE_FALLBACK_MODEL = 'deepseek-v4-flash:cloud';
+const GOOGLE_FALLBACK_MODEL = 'gemma4:31b-cloud';
 
 const LEGACY_BLOCKED_MODELS = new Set(['codex-davinci-002']);
 const DEFAULT_OPENAI_MODEL = 'gpt-4o';