Restore per-NPC LLM backend selection (regressed in SDK extraction #225) #236

What this is

Restore per-NPC LLM backend selection, which regressed when openrange-pack-sdk was extracted (#225).

Before the extraction, cyber webapp NPC factories took a model: config key and constructed a concrete agent backend. That couldn't survive the strict pack-cannot-import-openrange boundary (it needed to build a concrete StrandsAgentBackend), so it was removed. Today every NPC in an episode shares the single RunConfig.npc_agent_backend.

Proposed shape

Mirror the entry-point pattern so pack code stays free of openrange imports:

• Add a backend_id: config key on NPC factories (a plain string the pack can carry).
• Add RunConfig.npc_backend_factory: Callable[[str], AgentBackend] | None that the runtime resolves at NPC start, mapping backend_id → concrete AgentBackend.
• The pack never imports openrange; the host wires the factory.

Where

• NPC base / AgentNPC in openrange_pack_sdk (packages/openrange-pack-sdk/src/openrange_pack_sdk/_protocols.py).
• NPC factories in packs/cyber_webapp/cyber_webapp/npcs/.
• RunConfig / NPC start in src/openrange/core/.

Acceptance

• A single episode can run two NPCs on two different backends, selected by backend_id.
• No openrange import appears in pack code.
• Integration test exercising two distinct backends in one episode.

This is the 🚧 Core item in ROADMAP.md that had no tracking issue.