docs: add AI productivity research, trust calibration, and exploration workflow

## New Content

### Trust & Verification (ultimate-guide.md)
- Section 1.7 "Trust Calibration: When and How Much to Verify" (~155 lines)
  - Research-backed stats (ACM, Veracode, CodeRabbit, Cortex.io)
  - Verification spectrum by code type
  - Solo vs Team strategies with workflow diagrams
  - "Prove It Works" checklist
- New pitfall: "Trust AI output without proportional verification"
- CLAUDE.md size guideline: 4-8KB optimal, >16K degrades coherence

### AI Productivity (learning-with-ai.md)
- Section "The Reality of AI Productivity" (~55 lines)
  - Productivity curve phases (Wow Effect → Targeted Gains → Plateau)
  - High-gain vs low/negative-gain task categorization
  - Team success factors
- Productivity trajectory table by pattern (Dependent/Avoidant/Augmented)
- 5 new sources (GitHub, McKinsey, Stack Overflow, Uplevel, DORA)

### Session Limits (architecture.md)
- "Session Degradation Limits" section
  - Turn limits (15-25), token thresholds (80-100K)
  - Success rates by scope (1-3 files: ~85%, 8+ files: ~40%)

### Exploration Workflow
- NEW: guide/workflows/exploration-workflow.md
  - Anti-anchoring prompts, 3-5 approaches pattern
- iterative-refinement.md: Script Generation Workflow (3-7 iteration pattern)
- anchor-catalog.md: Anti-Anchoring Techniques, Exploration/Iteration Prompts

### Reference Updates
- adoption-approaches.md: Empirical data section
- reference.yaml: New deep_dive entries, updated line numbers

Sources: MetalBear engineering blog, arXiv studies, Addy Osmani (Jan 2026)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

guide/architecture.md

@@ -269,6 +269,33 @@ When context usage exceeds a threshold, Claude Code automatically summarizes old
 | Specific reads | Know what you need | Read exact files, not directories |
 | CLAUDE.md | Persistent context | Store conventions in memory files |
 
+### Session Degradation Limits
+
+**Confidence**: 70% (Tier 2 - Practitioner studies, arXiv research)
+
+Claude Code's effectiveness degrades predictably under certain conditions:
+
+| Condition | Observed Threshold | Symptom |
+|-----------|-------------------|---------|
+| Conversation turns | **15-25 turns** | Loses track of earlier constraints |
+| Token accumulation | **80-100K tokens** | Ignores requirements stated early in session |
+| Problem scope | **>5 files simultaneously** | Inconsistent changes, missed files |
+
+**Success rates by scope** (from practitioner studies):
+
+| Scope | Success Rate | Example |
+|-------|--------------|---------|
+| 1-3 files | ~85% | Fix bug in single module |
+| 4-7 files | ~60% | Refactor feature across components |
+| 8+ files | ~40% | Codebase-wide changes |
+
+**Mitigation strategies**:
+
+1. **Checkpoint prompts**: "Before continuing, recap the current requirements and constraints."
+2. **Session resets**: Start fresh for new tasks (`/clear`)
+3. **Scope tightly**: Break large tasks into focused sub-tasks
+4. **Use sub-agents**: Delegate exploration to `Task` tool to preserve main context
+
 ---
 
 ## 4. Sub-Agent Architecture