diagnose: fix-9-igpu-dgpu hardcodes HIP_VISIBLE_DEVICES=1, which is the iGPU on a desktop where the dGPU enumerates first

Omarchy 4.0.2 (Arch), kernel 7.1.9; RX 7900 XT (gfx1100) + Raphael iGPU (gfx1036). Built from main at a6fa76d plus #394 (on plain main the note below reads Detected gfx targets: [], see #393; the plan is the same).

Current behavior

$ rocm diagnose
#1 (fix-9-igpu-dgpu) [LIKELY score=65/100] iGPU enumerated alongside dGPU and destabilising the runtime
   plan: Pin the runtime to the discrete GPU with HIP_VISIBLE_DEVICES so the iGPU is hidden.
     $ # Then pin HIP to the dGPU (typically index 1 when an APU is index 0):
     $ export HIP_VISIBLE_DEVICES=1
   note: Detected gfx targets: ["gfx1100", "gfx1036"]. Discrete GPU(s): ["gfx1100"]; integrated APU(s): ["gfx1036"]. ... do not assume the higher-numbered gfx target is the dGPU ...
   verify after fix: HIP_VISIBLE_DEVICES=1 python -c "import torch; print(torch.cuda.device_count())"

The note names the discrete GPU; the plan and the verify line still assume it is index 1. Here it is index 0:

$ HIP_VISIBLE_DEVICES=1 llama-server --list-devices
  ROCm0: AMD Ryzen 5 7500X3D 6-Core Processor (20254 MiB ...)      <- the iGPU
$ HIP_VISIBLE_DEVICES=0 llama-server --list-devices
  ROCm0: AMD Radeon RX 7900 XT (20464 MiB ...)

Following the plan pins HIP, and everything on it, to the iGPU. This is the common desktop layout (Ryzen 7000 iGPU + Radeon card). rocm fix fix-9-igpu-dgpu without --device-index is print-only (#303), so the harm is the advice itself.

Expected behavior

The plan and verify line use the ordinal of the GPU the note identified as discrete (=0 here), or fall back to the <dGPU-index> placeholder fix.rs:252 already uses — never a literal 1.

Root cause

check_9_igpu_dgpu_collision (crates/rocm-core/src/diagnose.rs:915-990) has discrete_targets in hand for the note but renders the command and verify strings as literals with 1; the Windows branch hardcodes setx HIP_VISIBLE_DEVICES 1 likewise.

Possible solution

Map the discrete target to its position among the GPU agents and render that into both lines; placeholder when it cannot be identified. Unit test: a fixture with gfx1100 before gfx1036 must produce =0. Related: #149 and #382 touch fix-9's flag and wording but keep the literal; #267 adds visibility-mask helpers in serve that could be reused. Happy to take this after #394.

Start in crates/rocm-core/src/diagnose.rs, especially check_9_igpu_dgpu around lines 915-990, and inspect the existing placeholder usage in fix.rs:252. Trace how discrete_targets relates to GPU-agent ordering, then find or add the unit-test fixture described in the issue. Done when both plan and verification output use the identified discrete GPU index, with a placeholder when it cannot be identified, and the gfx1100-before-gfx1036 case produces =0.