Fix Windows probes in the dasllama bench harness

[artarasov]

llama-bench --ref never spawned on Windows: _popen goes through cmd /c, which strips the first and last quote of the command line, so the quoted binary path breaks. das_version probed bin/daslang, which cmd can't resolve, and wmic is gone on Win11 24H2+ so the ram/gpu fields came back empty. All three fixed in profile_common.das. Found while benching SmolLM2 on a 14700HX box.

[borisbat]

Verified all three fixes on a Windows 11 Pro 26200 (25H2) box, through the real popen path rather than from a shell:

Claim	Result
cmd can't resolve bin/daslang	confirmed — 'bin' is not recognized; bin\daslang → 0.6.3
cmd /c strips the first and last quote	confirmed — "…\where.exe" "cmd.exe" → 'C:\Windows\System32\where.exe" "cmd.exe' is not recognized
The sacrificial outer wrap fixes it	confirmed
The wrap is "safe either way"	confirmed on a plain command, on one with a 2>nul redirect, and on a single bare token

Lint and format are clean on the changed file. Diagnoses all look right. One thing I think needs another pass before this lands.

The virtual-adapter preference sits in the branch that doesn't run

The premise is that wmic is gone on 24H2+. It isn't gone on this box — 25H2, and wmic answers fine:

Name=Parsec Virtual Display Adapter

Name=NVIDIA GeForce RTX 5060 Ti

platform_gpu() returns the first Name= line and short-circuits (profile_common.das:483-488, unchanged by this PR), so this box records Parsec Virtual Display Adapter as its GPU — and the new prefer-physical logic below it never executes. As written the ordering fix only helps boxes where wmic is actually absent, which is not the same set as "24H2+" (the WMIC feature-on-demand is deprecated and off by default on clean installs, but survives an in-place upgrade).

Suggest hoisting the selection into a small helper and running it over both outputs, or dropping the wmic branch entirely and always going through CIM.

Smaller things

• "Virtual" is a narrow filter. It catches Parsec and VirtualBox, but not Microsoft Remote Display Adapter, Microsoft Basic Display Adapter, Microsoft Hyper-V Video, VMware SVGA 3D, or Citrix Indirect Display Adapter. Case-sensitive as well.

• The wmic miss is noisy. run_capture doesn't redirect stderr, so on a genuinely wmic-less box the harness prints 'wmic' is not recognized as an internal or external command, to the console twice in the middle of a bench run. 2>nul on both wmic commands handles it — and it survives the popen path (tested above).

• The shell-free path already exists in-tree. daslib/fio.das:648 run_and_capture(args, output, timeout) goes through popen_argv → CreateProcess directly, so neither the quote rule nor / vs \ applies. das_version() converts in one line and the new is_windows() branch disappears with it. run_capture can't move wholesale — it runs real shell pipelines like lspci | grep — and run_llama_bench would need its three callers to build argv, so the sacrificial wrap is a reasonable pragmatic fix there. Worth naming popen_argv in that comment so the next person doesn't re-derive the cmd /? rule.

• run_capture has no f != null guard (profile_common.das:307) while run_llama_bench does. Pre-existing, but this PR is specifically about probes failing to spawn.

Thanks for chasing this down — the cmd quote rule in particular is a trap we keep re-hitting.