[winml inspect] [P0] `winml inspect -m <hf_id>` takes 24 s end-to-end; first user-visible output silent for ~14 s

[zhenchaoni]

Severity: P1
Command: winml inspect

---

Category: Time-to-First-Output — first byte to user must arrive within 500 ms of entry.

Repro:

Measure-Command { uv run winml inspect -m microsoft/resnet-50 }

Actual: ~24.3 s wall time on warm cache. The first ~14 s after pressing Enter shows nothing on screen, then the panels render almost instantly.

Expected: Print a banner immediately on entry (Inspecting microsoft/resnet-50 …), show a rich.status spinner during HF metadata fetch. inspect should never download model weights — only config.json is needed. With cached config.json, drop to <2 s.

Why it matters: A 14-second silence makes the user assume the command hung and Ctrl-C; then they re-run, get the same silence, and the issue compounds. A spinner inside the first 500 ms removes the ambiguity.

[DingmaomaoBJTU]

Triage summary

• Component: Inspect
• Priority: P1

winml inspect -m <hf_id> takes 24 s end-to-end with 14 s of silence. Add a progress indicator and investigate the startup latency.