Use FastAPI-compatible ORJSON encoding across Gym servers

Parent workstream: #2999

Slowdown path

flowchart LR
    A[Typed handler result] --> B[FastAPI response contract]
    B --> C[Generic Python traversal]
    C --> D[Standard JSON encoding]
    D --> E[Large response bytes]

Current behavior

Agent and resources servers create ordinary FastAPI applications and register typed handlers without a shared fast JSON response path. See SimpleResponsesAPIAgent.setup_webserver and SimpleResourcesServer.setup_webserver.

The model-server-only _orjson_dispatch_response fast path turns a completed model directly into a Response. It does not provide an app-wide implementation that retains FastAPI's response-model validation, field filtering, aliases, custom serializers, injected status and headers, cookies, background tasks, ready Response objects, or OpenAPI behavior. See _orjson_dispatch_response.

Impact and provenance

Measured

The evidence report measured the parity-preserving app-wide prototype at 7.5 to 9.5 rollouts/s for a training-sized end-to-end body, a 27% increase. Its isolated agent/resources hop improved from 22 to 45 requests/s. These results came from the report's loopback full-rollout and hop rigs; they are not production-cluster measurements.

Expected

A shared encoder should reduce CPU time on non-streaming JSON responses across agents and resources servers. The gain will depend on payload size and concurrency, so no production throughput improvement is assumed.

Implementation constraints

• Preserve FastAPI's complete response contract and OpenAPI schemas; change the final byte encoder rather than bypassing response processing.
• Preserve ready Response and streaming/SSE pass-through behavior.
• Keep telemetry wrappers, session cookies, MCP-exposed routes, status codes, headers, and background tasks unchanged.
• Use the shared server setup so eligible agent and resources routes inherit the behavior without per-server forks.

Acceptance criteria

• Eligible non-streaming agent and resources responses use the fast encoder after normal FastAPI response processing.
• Contract tests cover response-model filtering, aliases, custom serializers, include/exclude options, status codes, headers, cookies, background tasks, ready Response objects, OpenAPI schemas, telemetry wrappers, and MCP routes.
• Streaming and non-JSON responses remain pass-through.
• Request-validation and error response shapes remain compatible.

Benchmark plan

Run alternating baseline/candidate loopback benchmarks for small control payloads and training-shaped responses at concurrency 1 and 32. Measure requests/s, p50/p99 latency, bytes, and CPU time for agent and resources hops, then repeat the full driver → agent → model → verifier rollout rig. Report hardware, dependency versions, payload construction, sample counts, and variance.

Related work/PRs

• Parent: #2999
• Model-server fast path: #2867

Start with SimpleResponsesAPIAgent.setup_webserver in nemo_gym/base_responses_api_agent.py, SimpleResourcesServer.setup_webserver in nemo_gym/base_resources_server.py, and _orjson_dispatch_response in nemo_gym/base_responses_api_model.py. Compare these entry points with the shared server setup, then use the contract tests and benchmark plan to verify FastAPI behavior, pass-through responses, telemetry and MCP routes, and the reported performance measurements.