feat(slime): add SlimeRunner as the primary Python entry point

[lliquid]

Goal

Replace the `train.sh` + `config.yaml` + env-vars combo with a single Python class. Users write:

```python
from agentcore_rl_toolkit.backends.slime import SlimeRunner

SlimeRunner(
exp_id="gsm8k-3b-smoke",
agent_runtime_arn="arn:aws:bedrock-agentcore:...",
s3_bucket="my-bucket",
model_dir="/path/to/Qwen2.5-3B-Instruct",
data_path="/path/to/gsm8k_tiny.jsonl",
model_type="qwen2.5-3B",
).train(num_rollout=10)
```

…instead of editing 121 lines of bash + 9 lines of YAML + exporting 4 env vars.

Design

`SlimeRunner` (`src/agentcore_rl_toolkit/backends/slime/runner.py`) — a dataclass with six required per-experiment fields (`exp_id`, `agent_runtime_arn`, `s3_bucket`, `model_dir`, `data_path`, `model_type`) and ~20 optional kwargs covering cluster, ACR/toolkit, hyperparameters, and wandb. An `extra_flags: list[str]` escape hatch passes any slime / Megatron-LM / SGLang CLI flag through verbatim.

`.train(num_rollout=1)` shells out to reproduce `train.sh` step-by-step:

1. `pkill -9 sglang` + `ray stop --force` to clear stale processes
2. `ray start --head`
3. Source `${slime_dir}/scripts/models/${model_type}.sh` to get `MODEL_ARGS`
4. Write toolkit config (ARN, bucket, exp_id, etc.) to a temp YAML
5. `ray job submit -- python3 ${slime_dir}/train.py ` with `--custom-config-path` pointing at the temp YAML

`init.py` now re-exports `SlimeRunner` as the only public symbol. `integration/` and `patches/` stay where they are — `integration.rollout.generate_rollout` is still passed to slime as a dotted-path string.

`train.sh` stays in the repo, unchanged. It's the low-level escape hatch for users who need to debug the raw slime CLI, and serves as the reference for what the Python class replicates.

Plan: `docs/roadmap/committed/slime-runner-entrypoint.md` (on the PR #59 branch).

End-to-end smoke test

Ran `python train.py` (which calls `SlimeRunner(...).train(num_rollout=10)`) on Qwen2.5-3B-Instruct + 64-row GSM8K, 8 × H100, `slimerl/slime:latest` container:

Rollout	raw_reward
0	0.252
1	0.355
2	0.344
3	0.488
4	0.377
5	0.594
6	0.533
7	0.676
8	0.574
9	0.675

Reward climbs 0.25 → 0.68 — matches PR #61's `train.sh` smoke (0.27 → 0.63), confirming the Python entry point produces the same training dynamics as bash. All 10 train steps logged `train/loss`, `train/grad_norm`, `train/step` and progressed monotonically. No session failures; only Ray atexit teardown tracebacks (cosmetic).

Test plan

• Unit tests (`tests/test_slime_runner.py`, 6 tests) cover:
  • `_build_runtime_env` required keys (PYTHONPATH, CUDA_DEVICE_MAX_CONNECTIONS)
  • Wandb env forwarding (opt-in only when key is set)
  • `from_yaml` round-trip
  • Key kwargs reach the slime CLI (`--num-rollout`, `--lr`, `--tensor-model-parallel-size`, `--rollout-batch-size`, plus load-bearing integration dotted paths)
  • `extra_flags` escape hatch appends verbatim
  • Toolkit config YAML carries ACR pointers
• Full test suite passes (81 tests).
• End-to-end smoke (10-rollout training on Qwen2.5-3B, rewards climb, loss moves).

🤖 Generated with Claude Code