Improve extraction prompts and logic

[kaminoguo]

Files

File	Purpose
src/sqrl/agents/user_scanner.py	Stage 1: Flash model scans user messages for patterns
src/sqrl/agents/memory_extractor.py	Stage 2: Pro model extracts memories from flagged messages
src/sqrl/agents/project_summarizer.py	Stage 0: Generates project context from README
src/sqrl/ipc/handlers.py	Orchestrates the pipeline
specs/PROMPTS.md	Prompt specifications (PROMPT-001, 002, 003)

Current Logic

Stage 0: ProjectSummarizer (cached)
    └── Reads README.md → 2-3 sentence summary
    
Stage 1: UserScanner (Flash)
    └── Input: all user messages
    └── Output: indices of messages with patterns (corrections, preferences, complaints)
    
Stage 2: MemoryExtractor (Pro) - runs per flagged message
    └── Input: trigger message + 3 AI turns context + project summary
    └── Output: user_styles[] + project_memories[]
    └── Filter: confidence > 0.8

Models

• Scanner: google/gemini-3-flash-preview (SQRL_FAST_MODEL)
• Extractor/Summarizer: google/gemini-3-pro-preview (SQRL_STRONG_MODEL)

Test Data

• test_data/ - sample episodes
• python -m sqrl extract <episode.json> - CLI for testing

What to Improve

• Prompt quality (reduce false positives)
• Extraction accuracy
• Skip criteria refinement