Verify inference-localhost-dev readiness for GPT-2 Small SAE steering

[mdheller]

Goal

Inspect and document the current readiness of Agent Machine's local inference path for Neuronpedia-compatible SAE steering, starting with GPT-2 Small.

Scope

• Inspect the current state of inference server configuration in agent-machine.
• Confirm whether MODEL_SOURCESET=gpt2-small.res-jb is loadable today.
• Confirm whether Gemma 2B / Gemma 9B sourcesets are available.
• Document what HuggingFace weights are required and whether they are freely accessible without gating.
• Document the expected local endpoint shape:
  • POST /steer
  • same payload shape as neuronpedia.org/api/steer
• If gaps exist, open follow-up issues scoped to closing them.

Output

Add a documentation file, e.g. docs/inference-local-steering.md, recording:

• which sourcesets are ready
• what weights are needed and where to get them
• what the local endpoint URL and payload shape are
• what Agent Machine commands bring the server up
• what is not yet ready and why

Constraints

This is an inspection and documentation task, not an implementation task.

Do not implement anything not already present.

Acceptance criteria

• Documentation file exists and records the inspection result.
• The doc distinguishes verified repo evidence from inferred or externally checked model availability.
• Follow-up issues are opened for any concrete readiness gaps.
• Native validation / CI passes.