fix voice auto-install chain — use canonical lmstudio-community GGUF ids and respect non-lmstudio providers

[atomantic]

Summary

The voice agent's auto-install chain in server/services/voice/bootstrap.js was firing on startup with model ids lms get can't fetch, producing:

🎙️  voice: qwen/qwen3.5-9b unavailable () — trying next
🎙️  voice: lmstudio-community/Qwen2.5-7B-Instruct-GGUF unavailable () — trying next
🎙️  voice: meta-llama/Llama-3.1-8B-Instruct unavailable () — trying next
🎙️  voice: exhausted install chain ... — set voice.llm.model explicitly in Settings

Two root causes:

1. Wrong / non-existent ids. qwen/qwen3.5-9b was never released (Qwen line went 2.5 → 3 → 3.6); meta-llama/Llama-3.1-8B-Instruct is a gated HF repo that lms get can't pull without auth; qwen/qwen3-4b-2507 is the MLX-only slug. Each attempt failed in ~100 ms with empty stderr, which the regex-filtered warning rendered as the literal ().
2. Ran for the wrong backend. ensureToolCapableModel / preloadModel always invoked lms get / lms load regardless of the configured voice provider, so an Ollama user saw lms failures for a CLI they never asked for.

Changes

• Replace the chain with the un-gated lmstudio-community/...-GGUF ids that lms get actually accepts — and that match the curated localLlmCatalog.js entries, so the "recommended installs" UI and the voice auto-installer agree on the same model set:
  • lmstudio-community/Qwen3-4B-Instruct-2507-GGUF (~2.6 GB, current non-thinking Qwen3, fast TTFT)
  • lmstudio-community/Llama-3.2-3B-Instruct-GGUF (~2 GB, smaller fallback)
  • lmstudio-community/Qwen2.5-7B-Instruct-GGUF (~4.7 GB, classic workhorse)
  • lmstudio-community/Meta-Llama-3.1-8B-Instruct-GGUF (~4.7 GB, Llama fallback)
• Gate ensureToolCapableModel / preloadModel on the effective voice provider — mirrors resolveLlmEndpoint's fallback (deleted / CLI-type / endpoint-less providers still resolve to LM Studio, so we still provision in those cases). Only skip when the configured provider resolves to a usable non-lmstudio api-type backend (e.g. a working Ollama).
• Rewrite the per-attempt warning to surface the last non-empty line of stderr (or stdout) instead of the trailing newline — so failures show what went wrong instead of ().
• Extract isEffectiveLmStudioVoiceProvider helper near LMS_BASE.

Test plan

• cd server && npm test -- --run services/voice/ → 387 voice tests pass
• Restart PortOS with voice.llm.provider='lmstudio', voice.llm.tools.enabled=true, voice.llm.model='auto' — verify the install chain attempts a real id and either succeeds or logs an actionable failure reason
• Restart PortOS with voice.llm.provider='ollama' — verify no lms get log lines appear

[Copilot]

Pull request overview

Fixes the voice agent's auto-install chain so it (a) targets canonical lmstudio-community/...-GGUF ids that lms get can actually fetch, (b) only runs when the effective voice provider resolves to LM Studio, and (c) surfaces actionable failure reasons instead of empty () warnings.

Changes:

• Replace the default tool-model chain with un-gated lmstudio-community/*-GGUF ids that match localLlmCatalog.js.
• Add isEffectiveLmStudioVoiceProvider and gate both ensureToolCapableModel and preloadModel on it, mirroring resolveLlmEndpoint's fallback rules.
• Improve the per-attempt failure warning to print the last non-empty line of stderr/stdout instead of a trailing newline.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.