test: add Megatron-backend variant of Gym token-capture recovery functional test

[lauradang]

What is uncovered

Both token-capture recovery functional tests added in #4129 source the vLLM-only base
tests/functional/grpo_async_gym_single_controller.sh:

• tests/functional/grpo_async_gym_single_controller_sibling_recovery.sh
• tests/functional/grpo_async_gym_single_controller_streaming_recovery.sh

The Megatron path (MInf capture -> data-plane checkpoint -> process restart -> ledger finalization)
has no end-to-end coverage; only unit tests exercise the Megatron stager/preparer.

Why it is blocked

_require_minf_capture_hooks() in nemo_rl/algorithms/single_controller_utils/setup.py
(called at line 1201, raises at line 965) raises NotImplementedError when the pinned
megatron-core lacks RequestPayloadStager / RequestPromptPreparer from NVIDIA/Megatron-LM
PR #7015. Until 3rdparty/Megatron-Bridge-workspace/Megatron-Bridge is bumped to a revision
that includes #7015, policy.generation.backend=megatron ++token_capture.enabled=true fails at setup.

Steps once #7015 is pinned

1. Add grpo_megatron_generation_gym_single_controller_sibling_recovery.sh (and a streaming twin)
   that set BASE_TEST to grpo_megatron_generation_gym_single_controller.sh (already
   policy.generation.backend=megatron + mcore_generation_config.expose_http_server=true
   via examples/nemo_gym/grpo_qwen3_0_6b_megatron_generation_single_controller.yaml).
2. Register them in tests/functional/L1_Functional_Tests_SingleController.sh next to the
   vLLM variants (lines 191/194); use run_test fast only if runtime fits the Lfast budget.

Split out from #4129 per review discussion.

🤖 Generated with Claude Code