Add LM Studio as a local model provider (and fix keyless-provider discovery)

[hacksics]

Adds lmstudio alongside ollama as a first-class local runtime, and fixes a pre-existing bug that affects Ollama today.

The bug (independent of LM Studio, worth reading first)

Connecting a keyless provider without editing the endpoint stores no profile at all, and _ollama_models() reads that profile to decide the runtime was opted into — so the card goes green and the picker stays empty. Anyone who accepted the default http://localhost:11434 and clicked Detect got no model discovery.

Two halves:

• useProviderSetup's Test-is-also-Save path skipped setProvider whenever a provider already reported configured, which keyless providers do from the first render — there is no key to be missing.
• Discovery treated a stored-but-empty profile as absent. Connecting on the default endpoint has nothing to write, so {} is the correct representation of "connected" and is None is the right test.

The feature

LM Studio previously worked only by pointing the OpenAI provider at localhost, which burns the single provider:openai profile (making local and hosted OpenAI mutually exclusive), demands a key it ignores, and gets no liveness gating.

The new descriptor follows the existing pattern exactly — keyless, endpoint prefilled to http://localhost:1234, liveness-gated, live model discovery. Notable details:

• Discovery prefers LM Studio's typed /api/v0/models over /v1/models: the stable route lists embedding models alongside chat models with nothing in the id to distinguish them. /api/v0 is beta, so any failure falls back to /v1 unfiltered.
• Connecting a local runtime selects a model it actually has. A single named recommended_model nearly always misses when the catalogue is whatever the user downloaded, so local providers fall back to the first discovered model — otherwise a working local server still reads "No model".
• The placeholder key is passed explicitly rather than via secrets, so a real OPENAI_API_KEY in the environment can never reach a local endpoint — the same rule _openai_compat already follows.
• LOCAL_PROVIDERS / LOCAL_DEFAULT_URLS factor the shared behaviour; _normalize_ollama_url keeps its signature and tests.

Verification

Against a real LM Studio 0.3.x holding qwen/qwen3.5-9b, google/gemma-3-4b and text-embedding-nomic-embed-text-v1.5:

• discovery returns the two LLMs and drops the embedder
• a live tool-calling completion through ProviderRouter returned get_weather{'city': 'Colombo'}
• the Settings flow end-to-end: card reads "✓ Running", the discovered model lands in the picker, and an already-working default is not stolen

902 passed (backend, +11 new), 68 passed (vitest), 154 passed (playwright), tsc --noEmit clean.

Maintainer calls I'd rather you make than me

• lmstudio.webp is the only raster in providers/logos/ (13.6 KB) and is LM Studio's own app icon from lmstudio.ai — every other mark there is from the MIT-licensed lobe-icons set. I noted the provenance difference in logos.ts. Because it carries its own background it renders edge-to-edge via FULL_BLEED_LOGOS rather than floating at 60% on the light plate. Happy to drop the commit entirely if you'd rather not vendor it; it falls back to the neutral "L" monogram with no other change.
• Capabilities stay conservative for lmstudio (vision=False, parallel_tool_calls=False), matching ollama — even though both test models are vision-capable. capabilities_for() is a pure function of the model string and can't see the live server.
• Picker rows show the raw id (lmstudio:qwen/qwen3.5-9b). matrix.py deliberately curates no local entries, since which models exist is a property of the user's machine.

Screenshots

(to add)

🤖 Generated with Claude Code