feat(slime): support arbitrary agent payload shapes in the training backend

[lliquid]

Goal

Let the slime training backend carry any agent payload shape from the JSONL dataset to the agent, so agents with different input schemas (GSM8K, AppWorld, migration, OfficeBench, future ones) can train through slime without bespoke integration changes.

Before this PR, `_sample_to_payload` hardcoded the GSM8K shape (`prompt` + `answer` top-level plus a nested `metadata` dict), which forced every other agent into workarounds. After this PR, the JSONL row's `metadata` dict is the agent payload verbatim — the data author owns the schema, the framework holds no opinion.

User-facing effect

Each agent declares its payload shape by choosing what keys go in `metadata`:

Agent	JSONL row shape	Agent reads
Math (GSM8K)	`{prompt, metadata: {prompt, answer}}`	`payload["prompt"]`, `payload["answer"]`
AppWorld	`{prompt, metadata: {task_id}}`	`payload["task_id"]`
Migration	`{prompt, metadata: {repo_uri, metadata_uri, ...}}`	`InvocationRequest(**payload)`
OfficeBench	`{prompt, metadata: {task_uri, testbed_uri}}`	`InvocationRequest(**payload)`

The top-level `prompt` field stays — it drives slime's tokenizer / length filter. The agent-visible payload comes entirely from `metadata`.

How

One-line replacement in `backends/slime/integration/rollout.py::_sample_to_payload`: return a shallow copy of `Sample.metadata`. All the old hardcoded-shape logic is deleted.

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

Breaking change

JSONLs in the old `{prompt, label}` shape produce empty agent payloads under the new rule. Regenerate with the updated SETUP.md data-prep snippet, which emits `{prompt, metadata: {prompt, answer}}`. The math agent's `rl_app.py` already reads `payload["prompt"]` / `payload["answer"]` and works without changes once the JSONL is regenerated.

End-to-end smoke test

Ran `bash train.sh` with `NUM_ROLLOUT=10` on Qwen2.5-3B-Instruct + 64-row GSM8K (`slimerl/slime:latest` container, 8 × H100):

Rollout	raw_reward
0	0.271
1	0.308
2	0.346
3	0.540
4	0.441
5	0.633
6	0.552
7	0.597
8	0.458
9	0.556

Reward climbs from 0.27 → 0.63 over 10 rollouts — GRPO is learning under the new contract. All 10 train steps logged `train/loss`, `train/grad_norm`, and progressed `train/step` monotonically. No Tracebacks or FAILED session statuses (one transient ACR ThrottlingException was caught and retried; two atexit tracebacks from Ray teardown at job exit are cosmetic).

Test plan

• Unit tests (`tests/test_slime_rollout_payload.py`) pin the contract: metadata-in-is-payload-out, empty/missing/None/non-dict metadata → `{}`, returned dict is a shallow copy, `sample.prompt` / `sample.label` are ignored.
• Existing test suite passes (79 tests).
• End-to-end smoke test (10-rollout training on Qwen2.5-3B, rewards climb, loss moves).

🤖 Generated with Claude Code