Parent epic: #2141

Phase 4: Triage + Cascade hybrid

Tasks

• Composite routing: triage picks initial tier, cascade catches misclassifications
• Config: fallback_strategy = "cascade" in [llm.complexity_routing]
• If triage routes to cheap model and cascade detects degenerate output → escalate as normal
• Triage verdict stored in turn metadata for cascade to reference (avoid re-triage on escalation)
• Consider LlmRoutingStrategy::TriageCascade composite variant vs runtime composition

Key files

• crates/zeph-llm/src/router/triage.rs
• crates/zeph-llm/src/router/cascade.rs
• crates/zeph-llm/src/router/mod.rs

Acceptance criteria

• Misclassified complex task (triage says Simple, output is degenerate) → cascade escalates
• Correctly classified simple task → no escalation, cheap model response returned
• Composite routing config validates that cascade providers match triage tier providers