fix(nemo-gym): survive a real multi-server NeMo-RL GRPO run

[Meirtz]

Context

Bringing up a real NeMo-RL GRPO run (not a bespoke trainer) over the NeMo Gym resources server backed by ShinkenComputerEngine fork-reset surfaced three integration breaks between our shinken.integrations.nemo_gym server and the current NeMo Gym multi-server agent. Each is fixed here with a failing-first test.

Fixes

1. Starlette lifespan compat — newer Starlette dropped add_event_handler, so the engine-shutdown binding crashed the server at startup with AttributeError. Fall back to wrapping the router's lifespan_context (same start-maintenance / close-engine lifetime contract).

2. Generation fence over a lossy multi-server transport — the per-rollout generation is fenced via the Starlette session cookie, but each server namespaces its own cookie (ShinkenComputerResourcesServer___shinken_cua vs SimpleAgent___...) and the agent chains resources_server_cookies from one tool response to the next. A tool/verify route that only read the session never re-emitted our Set-Cookie, so the chain broke and the next call arrived session-less (or carrying a sibling server's cookie) → the fence rejected it with "no rollout generation". Two-part fix:

   • recover the generation from the engine keyed by the reliably-threaded session_id when the cookie drops it (ShinkenComputerEngine.current_generation);
   • re-assert the generation into the session on every tool/verify response, so Starlette re-emits the cookie and the agent's chain keeps carrying it end to end.

3. SHINKEN_NG_DEBUG request tracer (off by default) — logs inbound cookies / session_id / generation per request; the diagnostic that pinned the cross-server cookie drop. Kept for adapting to other agents.

Verification

Closed a real NeMo-RL GRPO step on one GPU node (Qwen3-4B, vLLM generation + DTensor training, container-matched run_grpo_nemo_gym.py): the validation set and the num_prompts_per_step × num_generations_per_prompt training group were collected as forks of the golden desktop checkpoint, then Processing rewards → Computing advantages → Computing logprobs, clean exit. README §4 and status.md are reconciled from "config sketch, never run" to the verified run.

tests/test_nemo_gym_integration.py gains coverage for the lifespan shim, the session_id fallback, the cookie re-assert, and current_generation recovery — 51 pass; full SDK suite green.

Not in scope (follow-ups)

The ≥500-rollout scale run over the real CUA-Gym corpus (in flight), and generalizing the 16-way concurrency soak into an env-gated live test.

🤖 Generated with Claude Code