fix: support generic OpenAI-compatible local model servers (Ollama) in the guided install

[maxtechera]

What

create-answer-engine aborts the local (lmstudio-provider) install at stage 2/6 when the model server is a generic OpenAI-compatible server rather than LM Studio itself. prepareLmStudioModels probes LM Studio's native management API (GET <origin>/api/v1/models, POST <origin>/api/v1/models/load); servers like Ollama serve /v1/* (chat, embeddings, models) but 404 on the native endpoints, so the install dies with Could not inspect LM Studio model state: HTTP 404.

This PR treats a 404 from the native inspect endpoint as "no management API — models load on demand" and skips the pre-load step. Model availability is still validated against /v1/models earlier in resolveModelSetup, and a reachable LM Studio keeps the exact current behavior (inspect + explicit load).

Why

Found while doing the v1.1.1 guided install (INSTALL_AGENT.md) on Windows 11 WSL2 with Ollama as the local model runtime. Everything else in the local route already works with any OpenAI-compatible server:

• --lm-studio-url http://127.0.0.1:11434/v1 + --models chat=...,embedding=... validates fine against Ollama's /v1/models
• containerBaseUrl correctly rewrites localhost → host.docker.internal (verified reachable from containers)
• The runtime's OpenAiCompatibleProvider chat + embeddings calls work against Ollama (verified from inside a container: chat completion with qwen3:30b-a3b-instruct-2507-q4_K_M, embeddings with nomic-embed-text at 768 dims)

This one function was the only hard blocker.

Testing

• New unit test: 404 on the native endpoint → resolves without a load call (existing LM Studio test unchanged)
• packages/create: 149/149 tests pass, type-check and build clean
• Ran the built prepareLmStudioModels + listLmStudioModels against a real Ollama 0.20.2 server: lists models, gracefully skips pre-load

Extra setup notes from the same install attempt (not part of this diff)

1. The v1.1.1 guided install is currently impossible for everyone: the v1.1.1 (and v1.1.0) GitHub Releases are still drafts with zero assets — the runbook's pinned bootstrap URLs 404. The Installer release assets workflow has never been dispatched. Inputs that line up for it: tag=v1.1.1, source_commit=a346a8c6a7833596c088ccdebadd69e3608c386b, runtime_image=ghcr.io/the-answerai/answer-engine@sha256:418c93dac3753e0850f5ba6627581f888acdd65f3b91b655bd898cee56f10701. Worth confirming the built bootstrap matches the checksum pinned in the tagged INSTALL_AGENT.md (ea7e4632…) before publishing.
2. Ollama users need an embedding model pulled first (ollama pull nomic-embed-text, 768 dims) — Ollama typically ships with chat models only.
3. Cosmetic: preflight's MODEL_RUNTIME check probes the lms CLI only, so it warns "LM Studio was not detected" even when an OpenAI-compatible server is up (non-blocking; probing the configured /v1/models URL could fix it in a follow-up).
4. Model IDs must match /v1/models exactly — for Ollama that includes the tag (e.g. nomic-embed-text:latest).

🤖 Generated with Claude Code

Closes #93

[bradtaylorsf]

Temporarily closing and reopening to retrigger CI after correcting the SHA-pinned Actions allowlist; no code state is being changed.