chore: guard the v0 legacy bridge against non-renderer training clients

[mikasenghaas]

Summary

Hardens the v0 legacy bridge's training client builder (LegacyEnvServer._v0_client).

• _v0_client hardcoded a v0 renderer (token-in/out) client. That's correct today — prime-rl trains renderer-only (the orchestrator's student and teacher pools both pin train_client_type="renderer", and it isn't a user-config knob).
• But the v1 ClientConfig is a discriminated union (type="renderers" vs type="openai"/MITO). A MITO config reaching this path would have silently built a renderer client from an OpenAIClientConfig (its missing renderer/renderer_model_name fall back via getattr), doing TITO instead of the requested MITO rather than failing.

Now:

• Raise a clear ValueError if _v0_client gets a non-renderer config — MITO (chat-completions) training of v0 envs isn't supported, so fail loudly instead of running the wrong inference mode.
• With the renderer type now guaranteed, read renderer / renderer_model_name / pool_size directly instead of via defensive getattr.

This is the natural seam at which to add real type="openai" dispatch if MITO training of v0 envs is ever enabled.

Verification

• ruff check --isolated / ruff format --isolated --check clean.
• _v0_client raises ValueError on an OpenAIClientConfig; a RendererClientConfig flows through unchanged.

Note

Guard LegacyEnvServer._v0_client against non-renderer training clients

Adds a strict isinstance check in _v0_client that raises ValueError when client_config is not a RendererClientConfig. Also replaces getattr-based access with direct attribute reads for renderer_model_name, renderer, and pool_size.

^{Macroscope summarized aef8445.}

---

Note

Low Risk
Narrow guard in the legacy training path; renderer configs behave as before, with clearer failure for unsupported MITO training.

Overview
LegacyEnvServer._v0_client now requires a RendererClientConfig. If training passes an OpenAIClientConfig (MITO / chat-completions), it raises a ValueError instead of silently building a renderer client with missing fields and running token-in/token-out training.

After the type check, renderer, renderer_model_name, and pool_size are read directly on the config (no getattr fallbacks).

^{Reviewed by Cursor Bugbot for commit aef8445. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved

Small validation change that adds an explicit type guard to the v0 legacy bridge, failing fast with a clear error message when a non-renderer client is passed instead of silently using defaults. This is defensive hardening with limited scope.

^{You can customize Macroscope's approvability policy. Learn more.}