examine: gfx_target is empty for every GPU even though rocminfo lists them → ISA `Name:` lines overwrite the agent name

Reported from Omarchy 4.0.2 (Arch), kernel 7.1.9, outside the documented Ubuntu 24.04+ matrix. The defect is in a pure text parser and does not depend on the distro; reproduced with a binary built from main at a6fa76dbca5d767c1190ee47b765d16bad0e4a53.

Current behavior

rocm examine --json reports an empty gfx_target for every GPU even though rocminfo is present, runs cleanly (rocminfo_status: "ok"), and lists both GPUs with their gfx names:

$ rocm examine --json --framework llama-cpp | jq -c '.rocminfo_status, .summary.detected_gfx_target, (.gpus[] | {name: .name[:40], gfx_target, is_apu})'
"ok"
"gfx1100"
{"name":"Advanced Micro Devices, Inc. [AMD/ATI] Navi 3","gfx_target":"","is_apu":false}
{"name":"Advanced Micro Devices, Inc. [AMD/ATI] Raphael","gfx_target":"","is_apu":true}

summary.detected_gfx_target is right only because it comes from the separate sysfs path (detect_linux_sysfs_gfx_target); the per-GPU records never get a target from rocminfo.

Expected behavior

gpus[0].gfx_target == "gfx1100" (RX 7900 XT) and gpus[1].gfx_target == "gfx1036" (Raphael iGPU), matching what rocminfo prints, with is_apu derived from the target rather than guessed from the PCI name.

Steps to reproduce

Any Linux host with a working rocminfo (ROCm 6.x or 7.x):

rocm examine --json | jq '.rocminfo_status, [.gpus[].gfx_target]'
# "ok"
# ["", ""]

Root cause

probe_gpus_rocminfo (crates/rocm-core/src/examine.rs:1094-1116) walks rocminfo's output keeping a running cur_name, and every line that starts with Name: overwrites it. rocminfo prints the ISA sub-entries of a GPU agent after Device Type: GPU, and those lines also start with Name::

Agent 2
  Name:                    gfx1100
  Marketing Name:          AMD Radeon RX 7900 XT
  Device Type:             GPU
      Name:                    amdgcn-amd-amdhsa--gfx1100
      Name:                    amdgcn-amd-amdhsa--gfx11-generic
Agent 3
  Name:                    gfx1036
  ...

So when the next Agent header arrives, cur_name is amdgcn-amd-amdhsa--gfx11-generic, cur_name.starts_with("gfx") is false, and the agent is dropped. The same happens to the last agent at end of input. gfx_targets ends up empty, the function returns early, and no GPU record is touched. Re-implementing the loop verbatim over this host's rocminfo output confirms gfx_targets == []; accepting only the first Name: after an Agent header yields [("gfx1100", "AMD Radeon RX 7900 XT"), ("gfx1036", "AMD Ryzen 5 7500X3D 6-Core Processor")].

Every rocminfo since ROCm 5 prints ISA entries this way, so this affects every host where rocminfo is the source of the target.

Impact

• rocm diagnose filters on per-GPU gfx_target (crates/rocm-core/src/diagnose.rs:354, 943-950) to tell the discrete GPU from the APU. With every target empty, the fix-9-igpu-dgpu cause prints Detected gfx targets: [] and falls back to its generic advice instead of naming the discrete GPU (observed on this APU+dGPU host).
• Any consumer of examine --json that reads gpus[].gfx_target or gpus[].is_apu gets empty or PCI-name-guessed values.

Possible solution

Only accept a Name: line as the agent name when none has been recorded since the last Agent header (or ignore names that start with amdgcn-); commit the agent on the next header and at end of input as today. A unit test feeding the verbatim rocminfo excerpt above should assert two GPU targets. I can take this if nobody is on it.

Your environment

• rocm-cli: built from main at a6fa76d (rocm 0.1.0); also reproduced with the release-channel binary
• Platform: Linux, Omarchy 4.0.2 (Arch), kernel 7.1.9-arch1-2
• GPU/APU: AMD Radeon RX 7900 XT (gfx1100) + Ryzen 5 7500X3D iGPU (Raphael, gfx1036)
• rocminfo: ROCm 7.2.4 at /opt/rocm/bin/rocminfo (distro packages, left unmanaged by rocm-cli); managed runtime release-wheel-multi-arch-7-14-1 also installed, same result

Start in crates/rocm-core/src/examine.rs:1094-1116 at probe_gpus_rocminfo and trace how Agent and Name: lines are parsed. Add a unit test using the verbatim rocminfo excerpt from the issue, then verify it produces gfx1100 and gfx1036 GPU records with the expected APU classification.