feat: design hierarchical routing and recipe chaining without becoming a workflow engine

Summary

Design hierarchical routing and recipe chaining so one routing decision can delegate to another bounded routing policy, without turning vLLM Semantic Router into a general workflow engine.

Parent roadmap: #2287 Related: #1815, #2322, #2331, #2333

Motivation

Users have asked for hierarchical and multi-step routing: a top-level router picks a domain or agent, then a secondary router makes a more granular decision. This is related to, but different from, fixed multi-step model pipelines.

Scope

Define a bounded router-native model for:

• nested route decisions;
• recipe-to-recipe delegation;
• sub-router candidate constraints;
• recursion and depth limits;
• diagnostics for parent and child decisions;
• interaction with entrypoints and virtual model names.

Non-goals

• Do not implement arbitrary DAG orchestration.
• Do not replace external workflow engines.
• Do not mix this with fixed multi-step pipelines; #1815 tracks that use case.

Acceptance criteria

• A design clearly separates hierarchical routing from workflow orchestration.
• Delegation depth, cost, and recursion behavior are bounded.
• Replay can explain parent and child routing decisions.
• The design states whether implementation belongs in recipes, decisions, algorithms, or a separate layer.