feat: diagnose missing render nodes and surface SYCL crash errors

[offbyonebit]

$(cat <<'EOF'

Summary

• detect.py: Adds a note to DetectedGPU when a driver (xe/i915) is bound but no DRM render node appears in sysfs — the direct cause of SYCL's "No device of requested type available" crash. Exposes a new render_nodes_in_dev() helper that checks /dev/dri/ independently of sysfs.

• launcher.py: When no log_dir is provided, stderr is now captured to a tempfile instead of discarded to /dev/null. wait_ready() calls _surface_crash_logs() on early exit, which reads the tail of that file, logs it, and — if the canonical SYCL no-device message is found — emits a numbered checklist (render nodes, render group, dmesg, sycl-ls). The temp file is deleted in stop().

• cli.py: arc-llama doctor gains three new sections:

  1. /dev/dri/ render nodes — lists all renderD* entries (or explains why the directory is missing, e.g. in a container)
  2. sycl-ls device enumeration — runs sycl-ls if available and shows its output, making it immediately obvious whether SYCL can see the GPU
  3. Targeted WARNING block — fires when a GPU has its driver bound but no render node, printing the exact steps to resolve the issue on bare metal and in containers

Test plan

• pytest passes (46 pass, 9 skipped — no regressions)
• arc-llama doctor on a host with xe loaded but no /dev/dri/renderD* shows the WARNING block and numbered checklist
• arc-llama doctor on a working system shows green render nodes and sycl-ls device list
• llama-server crash with "No device of requested type available" now logs the checklist to the arc-llama logger

https://claude.ai/code/session_01GNwxSoERiDh1WWnS7Z2LvM
EOF
)

---

Generated by Claude Code