Add Dynamo renderer transport

[AmeenP]

Summary

Adds complete Dynamo renderer + TITO nvext support to verifiers in one PR.

• adds ClientConfig.renderer_transport, defaulting to vllm, with public values vllm and dynamo
• wires RendererClient to the standalone renderers transport API and forwards vllm / dynamo
• pins the standalone PrimeIntellect-ai/renderers repo to renderers PR Question: Completion IDs for multi-turn conversations. #11 at 17005dd
• teaches OpenAIChatCompletionsTokenClient to use renderer_transport="dynamo"
• sends Dynamo TITO token-in requests through /chat/completions with nvext.token_data
• requests nvext.extra_fields=["engine_data"] and grafts nvext.engine_data.{prompt_token_ids,completion_token_ids} onto the standard response token fields
• keeps renderer-derived root stop_token_ids for Dynamo compatibility
• preserves caller-provided nvext.extra_fields while adding engine_data
• fixes prompt-id prefix matching for vf.Message input and forwards chat_template_kwargs to bridge tokenization
• accepts current renderers RenderedTokens bridge return values

Context

This now absorbs the follow-up from #1313 so the verifiers Dynamo work is complete in this PR.

Paired stack:

• PrimeIntellect-ai/renderers#11
• ai-dynamo/dynamo#9509
• prime-rl #2420 / #2446, which set renderer_transport="dynamo" for Dynamo renderer and TITO clients

Validation

• uv run ruff check verifiers/types.py verifiers/clients/renderer_client.py verifiers/clients/openai_chat_completions_token_client.py tests/test_renderer_client.py tests/test_openai_chat_completions_token_client.py
• uv run pytest tests/test_renderer_client.py tests/test_openai_chat_completions_token_client.py
• pre-push hooks: ruff check, ruff format, Semgrep policy, ty (ci parity)

[macroscopeapp]

🟡 Medium clients/openai_chat_completions_token_client.py:324

When priority is present in extra_body, it gets added to both nvext["agent_hints"] (line 286) and also to the body directly via the passthrough logic (lines 324-329). Since priority is not in promotable and body only contains top-level keys, the passthrough filter allows priority through, causing the same value to appear twice in the request — once at nvext.agent_hints.priority and once at body["priority"]. This duplicates the field and may confuse the Dynamo server or cause validation errors.

Consider adding priority to the passthrough exclusion filter so it is only sent via nvext.agent_hints.

+        passthrough = {
+            k: v
+            for k, v in extra_body.items()
+            if k not in promotable and k != "priority" and v is not None and k not in body
+        }

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

In file verifiers/clients/openai_chat_completions_token_client.py around lines 324-329:

When `priority` is present in `extra_body`, it gets added to both `nvext["agent_hints"]` (line 286) and also to the body directly via the passthrough logic (lines 324-329). Since `priority` is not in `promotable` and `body` only contains top-level keys, the passthrough filter allows `priority` through, causing the same value to appear twice in the request — once at `nvext.agent_hints.priority` and once at `body["priority"]`. This duplicates the field and may confuse the Dynamo server or cause validation errors.

Consider adding `priority` to the passthrough exclusion filter so it is only sent via `nvext.agent_hints`.

Evidence trail:
verifiers/clients/openai_chat_completions_token_client.py lines 281-329 at REVIEWED_COMMIT. Line 285 uses `extra_body.get('priority')` (not pop), leaving priority in extra_body. Lines 286-287 add it to nvext.agent_hints. Lines 324-329 passthrough filter does not exclude 'priority' (not in promotable tuple at lines 303-315, not a key in body dict built at lines 289-296), so it passes through to `body['priority']`.