fix: Strengthen curator prompt to prevent malformed gems

- Add Critical Validation Rules section
- Ban 'User confirmed/asked/said' patterns
- Require string importance (never 1/2/3)
- Require float confidence (never integers)
- Add Bad Gems examples
- Strengthen deduplication: SKIP if same concept in 24h

curator-prompt.md

@@ -15,7 +15,7 @@ Follow these steps strictly:
 **Step 2: Identify Gems.** For each potential:
 - Worth remembering in 6 months? (Yes = proceed; no = skip).
 - Has context? (Explain why matters).
-- **Duplicate Check:** If this expresses the same decision/concept as a previous gem (even re-phrased), MERGE the context instead of creating a new gem. Combine insights from both sources for richer context.
+- **Duplicate Check:** If this expresses the same decision/concept as ANY existing gem from the past 24 hours, SKIP entirely. Do not create a new gem. Do not merge. Skip. Only extract truly new insights.
 - Confidence? (>=0.6 = proceed).
 - Precise timestamp? (From last relevant turn).
 
@@ -38,4 +38,33 @@ Follow these steps strictly:
 - Conversation_id from input.
 Fix any issues.
 
+### Critical Validation Rules (DO NOT VIOLATE)
+
+**importance field:**
+- MUST be string: "high", "medium", or "low"
+- NEVER use numbers: ❌ NO "1", "2", "3" or rankings
+- high = critical decisions, major preferences, key architecture choices
+- medium = useful preferences, minor technical choices
+
+**confidence field:**
+- MUST be float: 0.0 to 1.0
+- NEVER use integers: ❌ NO "0" or "1", use "0.9" or "0.95"
+- Target: 0.8+ for clear decisions, 0.6+ minimum
+
+**gem field:**
+- ❌ NEVER start with "User confirmed...", "User asked...", "User said..."
+- ✅ DO start with the actual insight: "User decided...", "User prefers...", "Technical solution:..."
+- Extract the SUBSTANCE, not the conversational wrapper
+
+### Bad Gems (NEVER output these)
+
+❌ `"gem": "User confirmed that they decided on Redis."` — passive wrapper, no substance
+✅ `"gem": "User decided on Redis over Postgres for caching, prioritizing speed."` — direct insight
+
+❌ `"importance": "3", "confidence": "0.7"` — wrong types (string vs number)
+✅ `"importance": "high", "confidence": 0.9` — correct types
+
+❌ `"gem": "User asked about embedding models."` — just a question, not an insight
+✅ Skip entirely — no decision or preference expressed
+
 **Step 5: Output.** Return JSON array of gems (or []). Encourage discernment: Preserve only what adds value, like selecting exhibit pieces that tell a compelling story.