Add explicit model routing for subagent dispatch

[WdBlink]

Problem

OPC deliberately fans work out across multiple independent agents, but its dispatch prompts currently call the built-in general-purpose agent without choosing a model. Claude Code defaults an omitted subagent model to inherit, so running OPC from a premium main session can multiply premium-model usage across every discussion, build, review, test-design, and UX observer.

This is separate from OPC's flow topology: the multi-node and multi-role quality design should remain intact. The missing piece is an explicit model-routing policy between role selection and agent dispatch.

Proposed direction

• Add layered agentRouting config under the existing ~/.opc/config.json / repo .opc/config.json mechanism.
• Resolve a model for every dispatched node/role with a deterministic opc-harness model-route command.
• Use host-native model identifiers. For Claude Code, sensible defaults can route economy work to haiku, standard implementation/review to sonnet, and reserve inherit/opus for explicit premium approval.
• Allow users and future host adapters (including Codex) to replace tier model IDs without changing OPC's DAG or role definitions.
• Detect CLAUDE_CODE_SUBAGENT_MODEL, since it has higher precedence than per-invocation routing.
• Fail closed when no model can be resolved; never silently treat an unknown/inherited child as a cheap model.
• Show the selected tier/model in OPC's dispatch preview and completion accounting.

Non-goals

• Reducing OPC's node count, review-role count, discussion rounds, or UX observers.
• Making MiniMax, a local proxy, or any third-party provider the public default.
• Changing the root model selected by the user.

I can submit a focused PR with the resolver, tests, protocol documentation, and prompt updates if this direction looks reasonable.