fix: expose DDP model view to online rollout

[0z5a]

Summary

• add an explicit executor-owned inference-model boundary for online rollout
• pass DDP.module to InferenceScheduler while retaining the wrapped DDP model for training and gradient synchronization
• fail fast for distributed FSDP and torch.compile online-rollout configurations before model/scheduler construction
• document the ownership contract and check in reproducible L20 + InfraSWE evidence

Why

TrainContextBuilder currently passes the training model wrapper directly to inference and reads context.model.config. A real DistributedDataParallel wrapper does not expose that attribute, so online GRPO/DPO fails while constructing its scheduler even though each DDP rank already owns a replicated underlying module.

The executor now declares whether an in-process rollout is supported and returns the model view inference may own. This also prevents sharded FSDP parameters from entering an unsupported generation path by accident.

Validation

Local

• ruff format --check .: 167 files formatted
• ruff check . --select I: passed
• focused CPU/Gloo boundary tests: 6 passed, 1 deselected
• complete suite: 626 passed, 104 skipped in 12.57s

NVIDIA L20 / SM89

Environment: NVIDIA L20, driver 570.86.10, CUDA 12.8, PyTorch 2.11.0+cu128, Python 3.12.12.

• all enabled CUDA extensions rebuilt for sm_89
• focused boundary tests: 7 passed in 1.54s
• complete built suite: 730 passed in 52.14s
• real CUDA/NCCL DistributedDataParallel wrapper + InferenceScheduler: the executor's inference view produced exactly the same greedy token lists as the underlying replicated model

The CUDA DDP probe uses world_size=1. A two-rank GPU run was not performed because two unowned idle GPUs were unavailable; this PR makes no multirank throughput claim. Full evidence is in docs/benchmarks/ddp_rollout_l20_sm89.json.

InfraSWE before PR

InfraSWE commit 811bc775ed5b3a6ec853219245f3469f78818020 was run before opening this PR:

• Draft/system-path engine subsets: 53 passed
• native ProjectComparisonCell validation: passed
• diagnostic project-fit-system-path-v0.5.1: 92.16/100
• diagnostic BenchmarkTrust: 89.79/100
• official score: unresolved, as expected without a seal, five fresh-process replays, system traces, hidden probes, and verified manifest

The score is explicitly non-official and records the missing two-rank cell and absence of a performance-statistics claim. Inputs and rationale are in benchmarks/results/ddp_rollout_l20_sm89_infraswe_score.json.