feat(k8s): minimal kubectl launcher mirroring slurm

[samsja]

Summary

Adds a k8s launcher path that mirrors the existing slurm flow: same RLConfig, new [k8s] block, single Jinja template, one kubectl apply to submit.

The point of this PR is to share the design with the team for early feedback — scope is intentionally small.

How it works

uv run rl @ examples/reverse_text/rl.toml @ examples/reverse_text/k8s_rl.toml

1. rl_k8s() writes trainer.toml / orchestrator.toml / inference.toml to ./k8s-runs/<job>/configs/ (analog of write_subconfigs in slurm).
2. Renders templates/rl.k8s.yaml.j2 with those TOMLs inlined as a ConfigMap (data.trainer.toml: | block scalar).
3. Manifest contains: ConfigMap, PVC, three headless Services, three StatefulSets (trainer / inference / orchestrator). Pods mount the ConfigMap at /etc/prime-rl/configs/ and the PVC at /data.
4. kubectl apply -f <manifest>.

Trainer rendezvous via <job>-trainer-0.<job>-trainer-headless:29500. Orchestrator gets INFER_URLS env from the headless inference service DNS pattern.

File map (5 changed + 1 gitignore)

file	change
src/prime_rl/configs/shared.py	new K8sConfig
src/prime_rl/configs/rl.py	k8s field, mutex with slurm, template auto-load, allow multi-node + k8s
src/prime_rl/entrypoints/rl.py	rl_k8s(), write_k8s_manifest(), dispatch in rl()
src/prime_rl/templates/rl.k8s.yaml.j2	new template
examples/reverse_text/k8s_rl.toml	example user-facing config
.gitignore	ignore k8s-runs/

Scope cuts (call out for review)

• Multi-node only. Single-node still uses rl_local.
• One pod per inference replica. No vllm-router fanout, no disaggregated PD. That's where the slurm template grows complexity; deferring until the basic flow lands.
• No NCCL/IB tuning. Slurm template auto-detects via ibv_devinfo; on k8s we rely on the GPU operator. May need an env block later.
• Existing k8s/ Helm chart left untouched as the legacy hand-driven path. Plan is to delete once this lands and disagg/router are ported.
• No live cluster test. Validated locally via yaml.safe_load_all (8 docs render, ConfigMap contains rendered TOMLs, trainer args reference /etc/prime-rl/configs/trainer.toml, INFER_URLS resolves correctly). kubectl --dry-run=client would catch schema issues but isn't installed in my env.