feat(report-agent): complete hermes bridge integration, curator loop, multi-job cron, and e2e testing

- Reduce role_max_items from 20 to 2-3 to avoid duplicate items
- Simplify deduplicate logic to take top N by score
- Add outfit_pairing_data.json with 59 products and 250 pairings
- Add outfit-logic UI visualizer for fashion rules
- Fix main_router.py indentation for pairing_router

Includes:
- OPENAI_API_KEY for zunef gateway (OpenAI-compatible)
- DEFAULT_MODEL=claude-sonnet-4-6
- Optional ANTHROPIC_API_KEY for direct Anthropic API
- Other services (Redis, StarRocks, Langfuse)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Inject English reminder message right before LLM call to force:
- All SKUs in response MUST be in product_ids
- Combo/set queries MUST return 5+ products
- Pick from suggest_items/ai_matches if tool returns <5
- Final self-check for (XXXXX) patterns
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add strict rules before JSON output:
- Any SKU in ai_response MUST be in product_ids
- For outfits: MUST include both Top + Bottom SKUs
- If tool returns <5 SP, MUST pick from suggest_items/ai_matches
- Self-check: scan ai_response for (XXXXX) before returning
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Update discovery section to:
- Require 5-8 products (up from 3-5)
- Force TOP + BOTTOM for combo/set queries
- Encourage picking from suggest_items/ai_matches if tool returns <5
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Move VD1 (combo di bien) to first example so Classifier
sees it before Early Exit section. Helps LLM trigger tool
for combo/set do queries.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Force tool_name=lead_search_tool for:
- combo/set do/phoi do queries
- sale/giam gia/discount queries (set discount_min=1)
- moi nhat/new arrivals (set discovery_mode=new)
- best seller queries (set discovery_mode=best_seller)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Prompt contains JSON examples with { } that confused .format().
Escape all braces first, then substitute only stage_injection.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Cherry-pick commit 88341cc4 with conflict resolution in graph.py.
Replace with_structured_output with manual JSON parsing for Claude API.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- Added hard_patterns.py: detect patterns (cheaper_than, more_expensive_than, similar_to) and extract reference product
- product_search_engine.py: integrate pattern detection, auto-set price_max/min based on reference price
- lead_search_tool.py: added user_insight parameter
- graph.py: pass user_insight to tool, removed product_ids filter (always return full tool result)
- prompts.py: added family/group intent examples, emphasize product_ids must include all items
- stylist_engine.py: added occasion and rule_id to matches, save to DB properly
- Added scripts: regenerate_matches.py and test_insert_one.py

Impact:
- Query 'rẻ hơn' auto sets price_max = ref_price * 0.95
- Query 'đắt hơn' auto sets price_min = ref_price * 1.05
- UI receives complete product list from tool (not truncated by AI)

- Removed orphan folder 08-lead-search/ (replaced by 08a/08b/08c/08d)
- Removed 8 orphan HTML/MD files from lead folders
- Fixed duplicate section titles (JS h2 + MD heading)
- Progress counter now auto-counts from registry.json
- Bumped cookbook.js to v9

- Journey Shopping (5 Modes): SVG funnel + 6 sections (Overview, Modes, Stage Transition, Upsell, Personalization, Troubleshooting)
- Dual-Agent Flow: SVG architecture + 6 sections (Overview, Classifier, Stylist, State, Tracing, Troubleshooting)
- Split Query & Cascading: SVG 7-tier + 6 sections (Overview, Literal, Inferred, Cascading, Enrichment, Troubleshooting)
- Modification Guide: SVG file map + 6 sections (FileMap, Tags, Tone, Upsell, NewTool, Checklist)
- Fixed mermaid.run() API call in cookbook.js (nodes param instead of querySelectorAll string)
- Bumped cookbook.js version to v8

fix(agent): make Stylist proactive — stop asking questions, auto-build combos from suggest_items + ai_matches

feat(agent): add COMBO rules + examples to Stylist prompt — use suggest_items and ai_matches to build full outfits

feat(agent): add 6 few-shot examples to Classifier prompt for combo, correction, recall, and early exit scenarios

fix(agent): prevent classifier from inheriting narrow product_line (like underwear) when user asks for a combo/set

fix(agent): update RECALL rules to correctly route queries with gender/product corrections to search tool

- config.py += FREE_LLM_API_KEY, LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY
- sqlite_db.py: use SQLITE_DB_PATH from constants
- n8n_desc.py: import FREE_LLM_API_KEY from config
- faq_route.py: import FREE_LLM_API_KEY from config
- live_monitor_route.py: import Langfuse keys from config
- prompt_optimizer_route.py: update docstrings (remove hardcoded IPs)
- NEW: tests/conftest.py with centralized BASE_URL

- NEW: backend/common/constants.py (SQLITE_DB_PATH, SQLITE_GENDER_MAP, KID_SAFE_KEYWORDS, normalize_gender)
- NEW: config.py += CHATBOT_API_BASE, LANGFUSE_SELF_HOSTED_URL, N8N_BASE_URL, LITELLM_BASE_URL, STARROCKS_INGEST_HOST
- REFACTORED: stylist_engine.py (removed duplicate maps, inline imports, dead _get_fallback_mappings)
- REFACTORED: lead_search_tool.py, outfit_matcher_worker.py, sqlite_mock.py, outfit_db.py (use SQLITE_DB_PATH)
- REFACTORED: live_monitor_route, prompt_optimizer, user_simulator, simulation_runner (use CHATBOT_API_BASE)
- REFACTORED: faq_route.py, n8n_desc.py (use LITELLM_BASE_URL)
- ARCHIVED: seed_rules.py, migrate_outfit_matches.py, check_*, inspect_db.py, recover_links.py to _archive/
- ZERO hardcoded IPs remaining in runtime code