Add Mobile300 Prime-RL tail-window staging

[bingran-you]

Summary

• add bench train run sft --message-tail-truncation keep-first-user for local Prime-SFT JSONL/data-dir staging
• update the env0-mobile300-pr828 compatibility profile to keep the first user instruction plus the longest final suffix that fits Prime-RL's data.seq_len * data.micro_batch_size window
• record tail-window stats in train-run.json and add an optional train extra pinned to Prime-RL's transformers==5.6.2

Why

Prime-RL's SFT StackDataset/CatDataset truncates overlength examples from the head. The historical custom trainer kept the labeled tail. This wrapper change prevents Prime-RL from seeing overlength Mobile300 rows without patching Prime-RL.

Validation

• uv run pytest tests/test_train_cli.py -q -> 26 passed
• uv run ruff check src tests tools
• uv run ruff format --check src tests tools
• git diff --check
• uv run pytest -q -> 4703 passed, 14 skipped, 7 deselected
• no-spend Mobile300 staging dry-run with real Qwen/Qwen3.5 tokenizer: 300 rows, 300 tool schemas omitted, 300 chat-template kwargs rows, 97 tail-windowed rows, max before 68105, max after 8188, max area 8192