* Studio: keep staged model's load settings visible while it loads
* Studio: pin staged load settings at click time, match loading pick by variant
* Studio: drop any stale staged pick after a successful load
* Studio: disable staged Load button while a different model loads
When a model is staged with "Load on selection" off and a different model (or a different GGUF variant of the same repo) is already loading, selectModel's in-flight-load guard matches on id + native path token only, so the click was silently deduped to a no-op. The staged Load button stayed enabled but did nothing, and the stale stage was then cleared once the other load finished.
Disable the staged Load button (showing "Another model loading...") whenever a different model is loading, so it is no longer an enabled no-op. The stage stays put and the user can retry once the in-flight load settles.
* Studio: refuse loading a different model while one is in flight
The in-flight-load guard in selectModel matched on id + native path token only, so a different GGUF variant of the same repo fell through and was silently deduped to a no-op. The earlier commit disabled the staged Load button for this case, but other entry points (e.g. selecting a different quant from the model picker with "Load on selection" on) still hit selectModel directly and no-op'd.
Make the guard variant-aware (id + GGUF variant + native path token) and, for a genuinely different model while a load is in flight, surface a "Another model is already loading" toast instead of silently returning. The load path has no clean supersession, so a second concurrent load is not started; the user is told to wait or cancel. Centralized in selectModel so every entry point is covered.
* Studio: lock staged model's load settings while it loads
The staged Load button snapshots context length, KV cache dtype, speculative
decoding, draft tokens, and tensor parallelism at click time, but the settings
sheet stays mounted during the load, so edits made while "Loading..." shows were
silently ignored and then overwritten by the load response. Disable those
controls while the staged pick is loading so the visible state matches what the
run actually uses.
* Studio: refuse to stage a model while another load is in flight
With Load on selection off, selecting a model mid-load staged it into
pendingSelection and showed a disabled "Another model loading..." button,
implying the user could retry once the load settled. The post-load cleanup then
treated that queued pick as stale and silently cancelled and cleared it. Refuse
to stage while a load (or a cancel's background unload) is in flight: stageModel
no-ops in that state so every entry point (the chat selector and the Hub) is
covered, and stageOrLoad surfaces a toast on the common path. The immediate-load
path is unchanged; selectModel already rejects a concurrent load.
* Tighten staged-load guard comments
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
49d4c61623