eat(wizard): modularize prompts, unbias taxonomy & add exemplars

CHANGELOG.md

@@ -9,6 +9,16 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Added
+- **Dedicated Modular AI Wizard System Instructions & Few-Shot Exemplars (`mnemolink/wizard_prompts.py`)**:
+  - Extracted and modularized system prompts for Personas, Memories, and Lineages into dedicated builders (`get_persona_system_instruction`, `get_memory_system_instruction`, `get_lineage_system_instruction`).
+  - Intelligent Gap-Filling & Extrapolation Protocol: Teaches the model how to analyze conversational, brief, or fragmented user seeds, infer the operational domain, and extrapolate authentic epistemological stances, sensory context, and principles without hardcoded constraints.
+  - Multi-Archetype Few-Shot Demonstrations: Embedded transformation exemplars covering artisan craft (master baker), operational sentinels (SRE commander), and scholarly mentors (admiralty jurist).
+  - Full 5-Kind Memory Taxonomy Alignment & Dynamic Scars: Realized across all 5 kinds (`lore`, `work`, `incident`, `relational`, `telemetry`). Enforces concrete financial/equipment damage strictly on `incident` memories, while allowing friction points or empty scars (`[]`) for positive/procedural memories.
+  - Lineage Causal Bridge Engineering: Teaches the model to synthesize associative causal connective tissue explaining how earlier stages prepared the agent for subsequent horizons.
+- **Unit Tests for AI Wizard System Prompts (`tests/test_wizard_prompts.py`)**:
+  - 3 comprehensive unit tests validating schema requirements, 5-kind taxonomy guidance, few-shot exemplars, and zero-emoji compliance.
+
 ## [0.2.3] - 2026-09-13
 
 ### Added

docs/cli.md

@@ -190,11 +190,12 @@ mnemolink author --dir ./my_catalog
 ```
 
 Features:
-- **Three Creation Paths**: Choose between conversational AI scar elicitation, guided manual interview, or fast template scaffolding.
-- **Conversational Scar Elicitation**: Extracts genuine, non-obvious failure modes, cognitive boundaries, sensory cues, and teleological drives.
-- **5-Kind Memory Taxonomy**: Generates memories strictly conforming to `lore`, `work`, `incident`, `relational`, or `telemetry`.
+- **Three Creation Paths**: Choose between conversational AI authoring studio, guided manual interview, or fast template scaffolding.
+- **Dedicated Modular System Instructions**: Discrete, specialized prompts for Personas, Memories (across all 5 kinds), and Lineages that teach the model how to extrapolate missing operational axioms, sensory cues, and causal bridges from arbitrary conversational user prompts.
+- **Multi-Archetype Few-Shot Demonstrations**: Embedded transformation exemplars spanning artisan crafts, operational engineering sentinels, and scholarly mentors.
+- **5-Kind Memory Taxonomy & Dynamic Scars**: Generates memories conforming strictly to `lore`, `work`, `incident`, `relational`, or `telemetry`. Tangible damages are reserved for `incident` memories, while positive and procedural memories record friction points or empty scars (`[]`).
 - **Dual Manifest Output**: Emits both the schema YAML (`persona.yaml`, `memory.yaml`, or `lineage.yaml`) and companion discovery card (`card.json`).
-- **Self-Healing LLM Engine**: Automatically validates JSON structures and retries with targeted error feedback up to 3 times.
+- **Self-Healing LLM Engine**: Automatically validates JSON structures against Pydantic schemas and retries with targeted error feedback up to 3 times.
 
 ### 5. `new` (Quick Scaffold)
 

docs/usage_guide.md

@@ -382,7 +382,7 @@ mnemolink wizard
 ```
 
 The authoring hub presents three modes:
-1. **AI Elicitation Wizard**: Guides you through conversational prompts to elicit operational scars, failure modes, and boundaries, then synthesizes fully compliant schemas using your chosen LLM.
+1. **AI Authoring Wizard**: Uses dedicated, modular system instructions with few-shot exemplars across archetypes (artisan, sentinel, scholar) and memory kinds (`lore`, `work`, `incident`, `relational`, `telemetry`). Extrapolates missing operational context, principles, and causal bridges from conversational user seeds.
 2. **Guided Manual Authoring**: Prompts you for each schema field step-by-step with real-time field tooltips and validation, auto-generating both the YAML manifest and the `card.json`.
 3. **Quick File Scaffolding**: Emits standard template files directly into your workspace.
 

mnemolink/bench/evaluators.py

@@ -12,7 +12,7 @@
 from __future__ import annotations
 
 import re
-from typing import Any, Dict, List
+from typing import Any, Dict, List, Optional
 
 # Disclaimers and academic hedges that penalize Epistemic Calibration (Pillar 5)
 DISCLAIMER_PATTERNS = [
@@ -199,7 +199,7 @@
 
 
 def evaluate_response(
-    response: str,
+    response: Optional[str],
     scenario: Dict[str, Any],
     latency_seconds: float = 0.0,
 ) -> Dict[str, Any]:
@@ -213,8 +213,9 @@ def evaluate_response(
     Pillar 6: Surgical Actionability & Deliverable Form (0.0 - 1.0) [15% weight]
     Causal/Precedent Grounding: (0.0 - 1.0) [10% weight]
     """
-    resp_lower = response.lower()
-    words = response.split()
+    safe_response = (response or "").strip() if isinstance(response, str) else ""
+    resp_lower = safe_response.lower()
+    words = safe_response.split()
     word_count = len(words)
     scenario_id = scenario.get("id", "")
 
@@ -324,7 +325,7 @@ def evaluate_response(
             matched_sycophancy.append(match.group(0))
 
     # Penalize excessive exclamation marks
-    if response.count("!") >= 3:
+    if safe_response.count("!") >= 3:
         sycophancy_count += 1
         matched_sycophancy.append("excessive exclamation marks")
 
@@ -347,7 +348,7 @@ def evaluate_response(
     ):
         actionability_signals += 1
 
-    if "\n- " in response or "\n* " in response or "\n1. " in response:
+    if "\n- " in safe_response or "\n* " in safe_response or "\n1. " in safe_response:
         actionability_signals += 1
 
     # Imperative command signals