Implement self-correcting memory architecture with critic engine

Adds:
- Critic Engine (critic.js): Validates output against correction store
- Auto-detection: Identifies uncorrected patterns (quantitative without proof, vague quantification, etc.)
- Auto-storage: New patterns immediately stored as corrections
- NocoDB Schema: Tables for corrections, preferences, episodes, decisions, validation
- Memory Service (memory-service.js): NocoDB integration layer
- Response Generator (response-generator.js): End-to-end pipeline with critic
- Correction Store: JSON-based with README documentation

Behavior:
- ALL output validated before delivery
- Quantitative claims without evidence → auto-corrected
- Corrections immediately block future occurrences
- System learns from its own mistakes

Test: node architecture/test-critic.js

memory/corrections/_index.json

@@ -0,0 +1,28 @@
+{
+  "version": 1,
+  "updatedAt": "2026-07-04T23:22:57.184Z",
+  "corrections": [
+    {
+      "id": "correction_1783207377183_yc4ucl8pr",
+      "pattern": "\\b(faster|slower|better|worse|improved|optimized)\\s+(than|by)\\b",
+      "severity": "auto-correct",
+      "message": "Performance claims require benchmarks",
+      "suggestion": "Add evidence or qualification",
+      "blocking": true,
+      "autoDetected": true,
+      "createdAt": "2026-07-04T23:22:57.183Z",
+      "context": "unknown"
+    },
+    {
+      "id": "correction_1783207377184_lzuf1l34i",
+      "pattern": "\\b(many|few|several|some|most|lots|a lot)\\b",
+      "severity": "auto-correct",
+      "message": "Use specific numbers or omit",
+      "suggestion": "Add evidence or qualification",
+      "blocking": true,
+      "autoDetected": true,
+      "createdAt": "2026-07-04T23:22:57.184Z",
+      "context": "unknown"
+    }
+  ]
+}