Rule 07: Add deprecation sweep for stale memory

[0SxD]

Source: Agents_Arcs audit. Scheduled scan of long_term.md for entries past staleness threshold without re-verification.

[0SxD]

Corridor validation (Agents_Arcs harsh audit): Rule 03 confidence field = 'confidence illusion' (OpenAI research). Reward hacking causes models to report high confidence when wrong. Need: explicit definition of what qualifies as high confidence, explicit 'I dont know' permission, defined routing path for sub-threshold items.