Hardening: thread-aware decision extraction + mid-confidence review queue

[cursor]

Background

M-005: naive single-message extraction hallucinates or misses decisions in noisy Slack threads.

Problem Statement

Real Slack/GitHub threads span multiple messages; low-confidence extractions can pollute the graph if written directly.

Proposed Solution

• Extract from thread/context window, not only a single message
• Classification-first (is this a decision?) then field extraction
• Confidence bands: >0.7 write; 0.4–0.7 human review queue; <0.4 discard
• Review UI or API for mid-confidence items (quarantine store may already help)

Acceptance Criteria

• Thread/context extraction path with tests
• Mid-confidence items go to review/quarantine, not silent graph write
• Documented thresholds in README or docs
• Fixture set of multi-message samples covered by tests

Files / Modules Affected

extraction/decision_extractor.py, pipeline/extraction_worker.py, memory/quarantine.py, frontend Review (optional), tests

Related: M-005 in MISTAKES.md