[RL-Kernel] Add user-facing strict, audit, auto, and off modes

[inaniloquentee]

Summary

This updates Vime's existing Qwen3-8B TP=2, CP=2 RL-Kernel integration
launcher and provider documentation. The PR targets main.

• adds RL_KERNEL_MODE=strict|audit|auto|off to the integration launcher;
• maps strict and audit to complete R/R routes;
• keeps auto compatibility-oriented with observable native fallback;
• makes off a clean P/P baseline with no selected-logprob provider, custom
  Megatron init, or vLLM plugin injection;
• gives strict and off the same aligned framework settings by default for
  a fair post-training E2E comparison;
• switches the documented provider path to
  rl_engine.integrations.vime.linear_logp.provider and explains the
  selected-logprob interface versus linear_logp operator naming.

This is intentionally limited to the integration script and documentation.
It does not modify Vime training behavior, Megatron integration internals, or
vLLM native code.

Tests

26 passed
bash -n scripts/run-qwen3-8B-rlkernel-tp2-cp2.sh
git diff --check

Companion RL-Kernel PR: RL-Align/RL-Kernel#344

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 26454ecf-afc3-44a7-aad6-93179c9f8cd2

---

_{Comment @coderabbitai help to get the list of available commands.}