Delegation-mesh auditing, agent-output exfiltration detection, coordination side-channel, MCP result and resource auditing

• A2A delegation-mesh auditing (mas-sentry a2a mesh). The single-target card
  audit reasons over one agent in isolation, but cross-agent weaknesses live on
  the delegation edges between agents, invisible to any one card - the
  overbroad-scope check shipped in 0.6.0 names cross-agent privilege escalation
  as its motive yet structurally cannot see it. The mesh scan takes an
  operator-declared topology ({agents: [{id, url}], edges: [[from, to]]}),
  fetches every card through the existing scope-enforced passive discovery, and
  builds a delegation graph carrying each agent OAuth2 scopes as node data.
  Topology is declared rather than inferred, mirroring --confirm-scope: the
  pentester maps the mesh they are authorised to test, instead of the scanner
  guessing edges from free-form card text (speculative) or observing them at
  runtime (needs authentication a passive scanner does not assume).
• Mesh detector: cross-agent privilege escalation through scope
  non-attenuation. Privilege attenuation, the 2026 A2A delegation consensus,
  requires every hop to carry equal or lesser authority than the hop before it.
  An edge A -> B where B advertises OAuth2 scopes absent from A is a
  non-attenuating hop: a task handed down it reaches authority A never held.
  Severity follows the contamination depth ladder - HIGH for a first-hop
  widening, CRITICAL when it sits two or more hops deep and compounds an

Full Changelog: v0.6.0...v0.7.0