Add Mobile300 Prime-RL token-suffix staging

[bingran-you]

Summary

Fixes the Mobile300 PR828 Prime-RL SFT compatibility path in BenchFlow by staging the dataset the way the historical custom trainer actually trained it:

• render each row with the historical custom-trainer chat-template path
• keep the exact rendered token suffix instead of message-level tail truncation
• omit tool_defs/tools for the compatibility profile, matching the archived PR828 rows
• use a generated passthrough Prime-RL chat template so Prime-RL trains on the staged suffix without rerendering tool schemas
• use assistant-only loss plus an unlabeled EOS pad message so stack packing gives one 8192-token row per sample

This keeps the fix in BenchFlow’s Prime-RL wrapper and avoids patching Prime-RL internals.

Why

The Mobile300 regression was not caused by the BenchFlow result JSONL schema alone. The key mismatch was that Prime-RL was tokenizing/truncating/loss-masking rows differently from the old custom trainer, causing a behavior collapse with many zero-tool eval attempts. The env0-mobile300-pr828 compatibility profile now encodes those historical semantics explicitly.

Validation

• uv run pytest tests/test_train_cli.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

Spendful validation is intentionally not in this PR; the next step is a real Mobile300 dry-run/train launch on the H100 from this branch.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 08934d5fc5

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]

Preserve per-row chat template kwargs when rendering

When a local JSONL row already carries chat_template_kwargs, this new suffix renderer calls apply_chat_template without forwarding them and _custom_trainer_token_suffix_row later drops the field from the staged row. In that scenario—for example Qwen rows that set enable_thinking=False—the staged token suffix is rendered with tokenizer defaults, so custom-trainer-token-suffix trains on a different token tail than the row/custom-trainer path requested.

Useful? React with 👍 / 👎.