Add comprehensive troubleshooting for Parts 18-21

Part 21: Real-Time Knowledge Sync — event-driven file watcher for LightRAG

Remove login page screenshots (looked ass)

Add live screenshots + image references in Parts 18 & 20

Apply Grok feedback: version line, Mermaid diagrams, real results

Fix inaccuracies using LightRAG knowledge graph for fact-checking

Add Parts 18-20: LightRAG graph RAG, Repowise codebase intelligence, agent observability + services stack

Major guide update: fix gaps that caused user setup failures

Added Ralph Wiggum Loop (force-continue until tests pass for autonomous work) + immutable vault entries (from Letta/MemGPT - write new, don't edit old, preserve decision trail)

Added context engineering metrics: 100:1 token ratio, compression targets (3:1 history, 10:1 tools), 60% utilization warning threshold from production data

Added harness engineering patterns: PreCompletion verification (LangChain +13.7pt), loop detection, multi-session progress files (Anthropic), exclusive file ownership (Zerg), worker self-verification. Updated one-shot prompt + checklist.

Part 15: Added secret leak prevention (3.2% GitGuardian stat, .gitignore template, commit scanning, secretgate proxy) + gateway crash-loop fix

Major guide update: Added coordinator protocol (4-phase from Claude Code), Cerebras + local models (QwOpus, TerpBot), autoDream in one-shot prompt + checklist, config protection, fixed Part numbering, updated templates with autoDream + coordinator + pure-index MEMORY.md, updated model recommendations

Part 16 v2: Updated with findings from actual Claude Code source files (autoDream.ts, consolidationPrompt.ts, memdir.ts). Added scan throttle, lock rollback, pure index format, grep-not-read approach, and hard caps from real implementation.

Part 16: autoDream Memory Consolidation — automatic memory cleanup inspired by Claude Code leak. Works with standard OpenClaw (no custom embeddings) or advanced vault setups.