M7: False positive learning from human decisions

[CodeBuildder]

Goal

When a human marks an alert as false positive, Argus remembers. Future identical alerts get higher false_positive confidence automatically.

Tasks

• Build fp_memory.py: store rule+pod+namespace+context hash → human decision
• Integrate into reasoning.py: if rule seen before and marked FP, boost false_positive confidence
• Add POST /incidents/:id/mark-fp and POST /incidents/:id/mark-real endpoints
• Build FP rate dashboard per rule
• Persist to disk

Acceptance criteria

• Mark alert as FP → next identical alert auto-classified as FP
• FP memory persists across agent restarts
• UI shows FP rate per rule
• memfd_create auto-classified as FP after 3 human confirmations