feat: add options.onlyLoaded to offer only loaded models

[dgnsrekt]

Closes #17.

Stacked on #48. That fix is a prerequisite — without it discovery returns nothing on current LM Studio builds, so this option cannot be exercised at all. Its commit is the first of the two here and will drop out of this diff once #48 merges. Review 1c40c7a alone for this feature.

Motivation

Discovery offers every generative record so LM Studio can load a model on demand when it is selected. That is the right default, but it has costs:

• With a large library the picker fills with models that are merely downloaded. Prune old LLMs from the list #17 describes exactly this — a list that accumulates everything present "at some point in time."
• On a shared or remote server, on-demand loading may be unwanted: picking an idle 400B model stalls the session while it loads, or evicts what someone else is using.

Change

provider.lmstudio.options.onlyLoaded restricts generated models and the generated whitelist to records with a non-empty loaded_instances:

{
  "provider": {
    "lmstudio": {
      "options": { "onlyLoaded": true }
    }
  }
}

Follows the options.* convention used for provider-scoped plugin settings (same shape as options.badge in #43).

Default behavior is unchanged — the option is opt-in, and only the boolean true enables it, so a malformed value can't silently empty the model list. When nothing is loaded the plugin generates an empty model list rather than falling back to idle entries; the discovery log gains onlyLoaded and skippedUnloaded so the filter is visible.

One incidental benefit: when a key has both a loaded and an unloaded record, the filter keeps the loaded one, so limit.context reflects the live allocation rather than whichever record landed last.

Verification

9 tests added (28 → 37): opt-in filtering, whitelist narrowing, default-unchanged, empty-when-nothing-loaded, explicit-override precedence, and a fail-open matrix over undefined | false | "true" | 1 | null.

npm run validate clean (lint, typecheck, 37 tests, build).

End-to-end against LM Studio 0.4.x, 37 records / 4 loaded instances, via opencode models lmstudio:

config	models offered
onlyLoaded absent	29
onlyLoaded: true	2

The 2 match the MODEL column of lms ps exactly (google/gemma-4-26b-a4b-qat, qwen/qwen3.8-27b).

Observation, not addressed here

That run also surfaced something pre-existing: my server reports the same key from more than one host, so loaded_instances for one key spans several machines and three native records collapse into two model entries via Object.fromEntries. It's why discovered counts records while the map holds fewer. I left the semantics exactly as they are on main rather than widen this PR — it looks related to #42, and I'm happy to open a separate issue with the payload if useful.

🤖 Generated with Claude Code

https://claude.ai/code/session_01S7Wjrr4wy4Eiig1PqtHfpD

[dgnsrekt]

Filed the duplicate-key observation from the PR description as #50, with a reproduction and the sanitized record shapes. It reproduces on main and is independent of this PR, so nothing here changes.

One thing that fell out of writing it up: onlyLoaded is immune to that collapse. The filter runs before the map is built, so when a key has both a loaded and an idle record the idle one is removed first and the loaded record survives — verified, not just reasoned about. That is the incidental benefit mentioned above, and it means this PR is unaffected either way #50 is resolved.