[Klaud Cold] Add /debug-mi300-enroot-pyxis command / 新增 /debug-mi300-enroot-pyxis 命令

[functionstackx]

Summary

• Adds .claude/commands/debug-mi300-enroot-pyxis.md, encoding today's MI300X debugging session: the enroot-nsenter: failed to create user namespace: Permission denied signature, root-on-controller access via ssh amd-vultr-mi300 with srun to reach compute nodes (no direct root SSH), a read-only survey of kernel.apparmor_restrict_unprivileged_userns across nodes, and the approval-gated fix (sysctl -w + /etc/sysctl.d/99-enroot-userns.conf) matching the cluster's working baseline.
• Documents the gotcha that plain unshare -U works even on broken nodes (Ubuntu's stock AppArmor profile), the rerun step, and the durable fix (bake the sysctl into the provisioning image — freshly provisioned nodes regress otherwise).
• Fix was validated live today: nodes chi-mi300x-043/049/057/058 were drifted at flag=1; after the fix, the three failed canaries ([Klaud Cold] Update qwen3.5-bf16-mi300x-sglang SGLang ROCm image to v0.5.14-rocm720-mi30x / 将 qwen3.5-bf16-mi300x-sglang 的 SGLang ROCm 镜像 升级至 v0.5.14-rocm720-mi30x #2062, [Klaud Cold] Update qwen3.5-fp8-mi300x-sglang SGLang ROCm image to v0.5.14-rocm720-mi30x / 将 qwen3.5-fp8-mi300x-sglang 的 SGLang ROCm 镜像 升级至 v0.5.14-rocm720-mi30x #2063, [Klaud Cold] Update kimik2.5-int4-mi300x-vllm vLLM ROCm image to v0.24.0 / 将 kimik2.5-int4-mi300x-vllm 的 vLLM ROCm 镜像 升级至 v0.24.0 #2069) were rerun.

中文说明

• 新增 .claude/commands/debug-mi300-enroot-pyxis.md，沉淀今日 MI300X 调试过程：enroot-nsenter 用户命名空间 Permission denied 的错误签名、通过 ssh amd-vultr-mi300（控制器 root）加 srun 访问计算节点（节点不接受直接 root SSH）、对 kernel.apparmor_restrict_unprivileged_userns 的只读全节点巡检，以及需用户批准的修复（sysctl -w + /etc/sysctl.d/99-enroot-userns.conf，对齐集群自身的正常基线）。
• 记录了易误导点：即使在故障节点上 unshare -U 也能成功（Ubuntu 自带 AppArmor profile 覆盖 unshare）；并包含重跑步骤与长期修复建议（将该 sysctl 写入节点供应镜像，否则新供应节点会回退）。
• 修复已于今日实测验证：chi-mi300x-043/049/057/058 四节点漂移为 flag=1；修复后已重跑三个失败的 canary（[Klaud Cold] Update qwen3.5-bf16-mi300x-sglang SGLang ROCm image to v0.5.14-rocm720-mi30x / 将 qwen3.5-bf16-mi300x-sglang 的 SGLang ROCm 镜像 升级至 v0.5.14-rocm720-mi30x #2062、[Klaud Cold] Update qwen3.5-fp8-mi300x-sglang SGLang ROCm image to v0.5.14-rocm720-mi30x / 将 qwen3.5-fp8-mi300x-sglang 的 SGLang ROCm 镜像 升级至 v0.5.14-rocm720-mi30x #2063、[Klaud Cold] Update kimik2.5-int4-mi300x-vllm vLLM ROCm image to v0.24.0 / 将 kimik2.5-int4-mi300x-vllm 的 vLLM ROCm 镜像 升级至 v0.24.0 #2069）。

🤖 Generated with Claude Code

[claude]

LGTM — documentation-only addition of a new debugging slash command.

Extended reasoning...

Overview

This PR adds a single new file, .claude/commands/debug-mi300-enroot-pyxis.md, which documents a debugging workflow for enroot/pyxis user-namespace failures on the MI300X Vultr cluster. It captures the error signature, controller-access pattern via ssh amd-vultr-mi300 + srun, a read-only sysctl survey, the approval-gated fix, and a note about the durable provisioning-image fix.

Security risks

None from this change itself — it is a Markdown runbook, not executable code. The fix it describes (setting kernel.apparmor_restrict_unprivileged_userns=0) does weaken a kernel mitigation, but the file appropriately gates that action behind explicit user approval (AskUserQuestion) and only aligns nodes to the cluster's existing working baseline. No secrets, credentials, or sensitive endpoints are introduced.

Level of scrutiny

Low. This is a .claude/commands/*.md slash-command reference file with no impact on production code, CI, or benchmark behavior. It only takes effect when a user explicitly invokes the command interactively.

Other factors

No bugs were found by the bug hunting system, there is no prior reviewer discussion to reconcile, and the runbook matches the incident described in the PR description (validated live on nodes chi-mi300x-043/049/057/058, with canaries #2062/#2063/#2069 rerun). Safe to auto-approve.