Define Neuronpedia-compatible local /steer endpoint contract

[mdheller]

Goal

Define the Agent Machine contract for a local Neuronpedia-compatible steering endpoint that Noetica can call without changing its client adapter.

Client-side contract references

This issue must align to the existing Noetica client shape rather than defining a parallel endpoint.

Reference Noetica files:

• SocioProphet/Noetica:lib/providers/neuronpedia.ts
• SocioProphet/Noetica:app/api/steer/route.ts
• SocioProphet/Noetica:docs/adapter-contracts.md
• Noetica PR Harden Agent Registry grant request scope #12: Make Neuronpedia base URL configurable

Noetica endpoint resolution after PR #12:

• NEURONPEDIA_BASE_URL=https://www.neuronpedia.org resolves to hosted /api/steer.
• NEURONPEDIA_BASE_URL=http://localhost:<port> resolves to local /steer.

Agent Machine should satisfy the localhost form.

Scope

• Define expected POST /steer request and response shape.
• Match Noetica's current payload shape:
  • prompt
  • model_id
  • steering.feature_id
  • steering.layer
  • steering.strength
  • optional steering.preset
• Define response shape compatible with Noetica SteeringResult:
  • status: applied | not_configured | noop
  • baseline
  • steered
  • diff_summary
  • feature_id
  • layer
  • strength
• Specify whether the endpoint is native Agent Machine, an inference-provider adapter, or an AgentPod workload.
• Specify health/readiness endpoint expectations.
• Specify how unavailable sourcesets return status: not_configured without crashing.
• Keep credentials and model weights out of config.

Acceptance criteria

• Contract or documentation exists for the local /steer endpoint.
• Contract explicitly references Noetica's client-side adapter and docs.
• Request payload includes prompt, model id, and steering config fields matching Noetica.
• Response payload supports Noetica's status: applied | not_configured | noop semantics.
• Contract states that MODEL_SOURCESET=gpt2-small.res-jb readiness is handled by Issue Register local steering sourcesets for GPT-2 Small and Gemma #33, not assumed here.
• No live provider implementation is required in this issue.

Non-goals

• Do not implement the server here unless a follow-up explicitly scopes implementation.
• Do not fetch model weights.
• Do not register sourcesets here; that is Issue Register local steering sourcesets for GPT-2 Small and Gemma #33.
• Do not add controlled activation here; that is Issue Implement controlled activation for local steering inference provider #34.
• Do not add production activation.