routing: make system prompts follow answer targets

[afourniernv]

Linear: SWITCH-1253

Problem

A client calls a Switchyard route without knowing which upstream model will answer. Those models may need different standing system instructions.

Today this is only configurable through Stage and Composite route fields such as efficient_system_prompt and capable_system_prompt. Other routers do not have the same configuration surface.

Fallback also matters. If a route selects fast and that target fails, the request sent to capable must receive the capable prompt. It cannot reuse a request already prepared for fast.

Proposed interface

Allow the prompt to be configured on a target:

[targets.fast]
id = "meta/llama-small"
llm_client = "together"
system_prompt = "Be concise."

[targets.capable]
id = "meta/llama-large"
llm_client = "together"
system_prompt = "Reason carefully."

[routes.agent]
id = "switchyard/agent"
type = "stage_router"
efficient_target = "fast"
capable_target = "capable"

The prompt should follow the answer target rather than the router type. Judge, classifier, and reviewer calls should remain unchanged.

Implementation

• Translation owns the model and prompt mutation for one request.
• libsy marks calls that may become the client-visible answer.
• libsy-llm-client prepares each selected or fallback candidate immediately before it is called.
• The native runner builds the effective prompt map from [targets.*].system_prompt.

This keeps provider request mutation out of libsy and prevents one candidate's prompt from leaking into a fallback.

Compatibility

• Existing configuration without target prompts behaves as before.
• Legacy Stage and Composite prompt fields remain supported.
• Target-level configuration takes precedence when both forms configure the same answer target.
• Routing decisions, retry behavior, and fallback order do not change.

Pull requests

• feat(translation): prepare requests for routed targets #455: request preparation helper, merged
• feat(llm-client): prepare routed completion candidates #463: answer-call boundary and candidate preparation
• feat(server): configure system prompts by target #464: target configuration, compatibility, tests, and docs

Done when

• Every built-in client-visible answer path uses the selected target's prompt.
• Each fallback is prepared independently.
• Existing caller instructions are preserved after the target prompt.
• Judge, classifier, and reviewer calls remain unprompted.
• Legacy configuration remains supported.