Part of #1195 — Phase 3

Implement quarantined LLM summarizer (Dual LLM pattern) for extracting facts from high-risk external content without exposing the main agent to raw untrusted data.

Crates: zeph-core
Depends on: SEC-1.4

Tasks:

• QuarantinedSummarizer struct wrapping Arc<dyn LlmProvider>
• System prompt: data extraction only, explicit instruction to ignore embedded directives
• Temperature 0, capped max_tokens, no tools available
• extract_facts(content: &SanitizedContent, prompt: &str) -> Result<String>
• Config:

  [security.content_isolation.quarantine]
  enabled = false
  sources = ["web_scrape", "a2a"]
  model = "haiku"
  max_tokens = 1024

• Wire into sanitization pipeline as optional post-processing step
• --init wizard: quarantine toggle and source selection
• Unit tests with mock provider, integration test with real content
• Metrics: quarantine invocation count, avg latency (via MetricsCollector)

Files: crates/zeph-core/src/sanitizer/quarantine.rs (new), crates/zeph-core/src/config/types.rs