[Yunnglin]

Features and Fixes

• Shared core HTTP multi-turn rollout, asynchronous entry points, and resource borrowing isolation.
• AgentENV GRPO cookbook, independent task scoring, and dedicated acceptance testing entry points.
• Fixed W1–W7: synchronous cancellation aggregation, initialization finalization, template paths, independent evaluation/seed, tool cancellation phase, wait-for-exit handling, and evidence determination.
• Preserve absent multimodal fields in text-only trajectories.
• Synchronize gradients on completed accumulation steps, fixing GA=2 updates that previously returned without changing weights.

Validation update — 9d918c0e

• Current-HEAD Transformers/Megatron × Twinkle/Tinker × SFT/DPO/GRPO: 12/12 passed, 0 skipped.
• SFT verifies LoRA + optimizer save/resume and continued training for both clients on both backends.
• Both backends confirm training instances STOPPED and training_resources_released=true; no validation processes remain.
• The GA=2 fix is also covered by a minimal LoRA update probe and 110 related regression tests.

Evidence:

• .qoder/evidence/d13-aeb73daf/matrix-9d918c0e-transformers/run.json
• .qoder/evidence/d13-aeb73daf/matrix-9d918c0e-megatron/run.json
• .qoder/evidence/d13-aeb73daf/scheduler-matrix-9d918c0e/scheduler.json

Pending

• Post-training sync/async parity investigation and the complete OpenEnv pipeline/learning/six-stage natural-cancellation suite are pending by user request.
• CI/lint cleanup is outside this round's scope by user request.
• OpenEnv functional coverage does not replace the AgentENV isolation/security gate.
• The 12/12 training matrix must not be interpreted as full D13 completion or completion of the pending items.