Add Prime-RL custom trainer compatibility controls

[bingran-you]

Summary

Adds explicit BenchFlow wrapper controls for reproducing the historical custom train_lora_sft.py data/rendering semantics while still launching native Prime-RL SFT:

• --renderer-mode none emits --renderer None, disabling Prime-RL's hand-coded renderer so SFT falls back to tokenizer apply_chat_template tokenization.
• --tool-defs-mode omit validates a local Prime-SFT source JSONL first, then removes tool_defs / tools only from BenchFlow's temporary training copy under <work-dir>/prime-rl-dataset.
• Records both choices in train-run.json so experiment artifacts preserve the exact compatibility semantics.
• Updates CLI docs and adds focused tests for the custom-trainer-compatible path and remote-dataset rejection.

Experiment impact

This targets the remaining env-0 Mobile300 trainer-swap mismatch: the archived custom trainer consumed rows with tool_defs but no tools, so it rendered without tool schemas and trained full rendered sequences. Prime-RL with [renderer] name="qwen3.5" instead reads tool_defs, inserts tool schemas, and masks via renderer sampled-token semantics. These flags let BenchFlow run Prime-RL with the custom trainer's effective input contract without modifying Prime-RL.

Validation

• uv run pytest tests/test_train_cli.py -q
• uv run pytest tests/trajectories/test_export_prime_sft.py -q
• uv run pytest tests/test_train_cli.py tests/trajectories/test_export_prime_sft.py -q
• uv run ruff check src/benchflow/training/backends/prime_rl.py src/benchflow/cli/train.py tests/test_train_cli.py
• uv run ruff format --check src/benchflow/training/backends/prime_rl.py src/benchflow/cli/train.py tests/test_train_cli.py
• uv run ty check src/benchflow/training/backends/prime_rl.py
• H100 Prime-RL parser check: cli(SFTConfig, args=[..., "--renderer", "None", "--dry-run"]) resolved renderer None.
• uv run bench train run sft --help

Spendful validation

No training rerun in this PR. The next env-0-experiment run should use:

TARGET_EXAMPLES=300 \
PACK_FUNCTION=stack \
LOSS_MASK=all \
MODEL_ATTN=sdpa \
RENDERER_MODE=none \
TOOL_DEFS_MODE=omit \
RUN_KIND=mobile300-custom-trainer-compatible-prime-rl \
bash experiments/env0-mobile-pr828-prime-rl-benchflow-cli/scripts/run_prime_rl_wandb_sft.sh

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8a0408cb15

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Avoid recursively copying the work dir into staged datasets

When --data points at a local dataset directory and the work dir is inside that directory (for example --data . with the default train-runs/sft), run_prime_rl_sft has already created <work-dir>/prime-rl before this copy runs. copytree then sees the new work directory under the source and recursively copies it into <work-dir>/prime-rl-dataset until it fails or consumes disk, so the documented local-dataset --tool-defs-mode omit path can break before launching training. Stage only the needed dataset files or ignore the work/staging subtree during the copy.

Useful? React with 👍 / 👎.