Clean routed experts response path

[S1ro1]

Summary

• route renderer generation through the PrimeRL generate endpoint without decoding routed_experts
• keep routed_experts as the compact opaque payload until prime-rl consumes it
• remove the prior docs/test churn from the routed-experts diff

Verification

• uv run --no-sync python -m compileall -q deps/verifiers/verifiers/clients/renderer_client.py deps/verifiers/verifiers/clients/openai_chat_completions_client.py deps/verifiers/verifiers/clients/openai_completions_client.py deps/verifiers/verifiers/utils/client_utils.py deps/verifiers/verifiers/utils/response_utils.py deps/verifiers/verifiers/types.py
• uv run --no-sync ruff check deps/verifiers/verifiers/clients/renderer_client.py deps/verifiers/verifiers/clients/openai_chat_completions_client.py deps/verifiers/verifiers/clients/openai_completions_client.py deps/verifiers/verifiers/utils/client_utils.py deps/verifiers/verifiers/utils/response_utils.py deps/verifiers/verifiers/types.py

Note

Normalize routed experts data in response parsing across client types

• Adds parse_routed_experts in response_utils.py to validate and normalize raw routed_experts dicts into a consistent RoutedExpertsPayload structure with typed data and integer shape fields.
• Applies parse_routed_experts in from_native_response for OpenAIChatCompletionsClient, OpenAICompletionsClient, and RendererClient so ResponseTokens.routed_experts always holds a parsed structure instead of raw model_extra values.
• Adds _generate_with_renderer to RendererClient as a new end-to-end generation path that handles renderer pool checkout, stop tokens, endpoint posting, and response parsing including routed_experts and tool calls.
• Tightens RoutedExpertsPayload.data from Any to str | bytes | bytearray | memoryview and enables arbitrary_types_allowed on CustomBaseModel.

📊 Macroscope summarized 28b092a. 5 files reviewed, 2 issues evaluated, 0 issues filtered, 1 comment posted

🗂️ Filtered Issues

[macroscopeapp]

🟡 Medium clients/renderer_client.py:741

The tool call ID assignment in from_native_response unconditionally overwrites the original id with f"call_{i}", discarding any ID provided by the model. If downstream code matches tool results to calls using the original ID, this breaks correlation.

-                    id=f\"call_{i}\",\n+                    id=tc.get(\"id\") or f\"call_{i}\",

🚀 Reply "fix it for me" or copy this AI Prompt for your agent:

In file verifiers/clients/renderer_client.py around line 741:

The tool call ID assignment in `from_native_response` unconditionally overwrites the original `id` with `f"call_{i}"`, discarding any ID provided by the model. If downstream code matches tool results to calls using the original ID, this breaks correlation.

Evidence trail:
renderers/base.py lines 496-502 (ParsedToolCall dataclass with `id` field), renderers/parsing.py lines 1111-1119 (Kimi K2 sets `id=raw_id or None`), verifiers/clients/renderer_client.py line 531 (`"tool_calls": parsed.tool_calls`), verifiers/clients/renderer_client.py lines 739-750 (`id=f"call_{i}"` unconditionally overwrites), verifiers/envs/tool_env.py lines 152-163 (downstream uses `tool_call.id` for correlation)

[S1ro1]

Closing this cleanup PR. The renderer-side fix is in renderers PR #54, so verifiers should stay on the normal renderers.client.generate path; prime-rl now pins renderers to that merged commit instead.

[S1ro1]

Reopening while we validate the renderers-only path against the routed-experts run. Do not merge or close yet.