feat(inference): auto-detect local providers in obol sell inference

[bussyjd]

What changed (pivot from review feedback)

Removed the first-run wizard that pushed users to sell inference on install. Instead, the auto-detection now lives inside obol sell inference — it runs only when the user explicitly chooses to sell and doesn't specify --model.

How it works

When obol sell inference <name> is run without --model in an interactive terminal:

1. Scans common local ports (8080 llama-server, 11434 Ollama, 8000 vLLM, 4000 LiteLLM)
2. Fingerprints server type via /api/tags, /health, /v1/models
3. Enumerates available models
4. Presents a selection:

Detected local inference servers:

  http://127.0.0.1:11434 (ollama) [healthy]
    - qwen3.5:9b (by library)
    - llama3.2:3b (by library)

  [1] qwen3.5:9b — http://127.0.0.1:11434 (ollama)
  [2] llama3.2:3b — http://127.0.0.1:11434 (ollama)

Select [1]: 

If --model is provided, detection is skipped entirely.

Files

File	Lines	Change
internal/inference/detect.go	~260	Keep: endpoint probing + FormatEndpointDisplay moved here from wizard
internal/inference/detect_test.go	~120	Keep: 6 tests unchanged
cmd/obol/sell.go	+40	Wire detection into sellInferenceCommand action
internal/inference/wizard.go	deleted	First-run wizard removed
internal/inference/wizard_test.go	deleted	Wizard tests removed

What was removed and why

The first-run wizard assumed users want to sell local inference on first install ("earn first, spend later"). The review feedback was clear: there's no PMF for that. The detection infrastructure is valuable — it just belongs in obol sell inference, not in a first-run flow.

Test plan

• go build ./cmd/obol/ compiles
• go test ./internal/inference/ — 6 detection tests pass
• Manual: run obol sell inference test without --model with Ollama running
• Manual: run obol sell inference test --model qwen3.5:9b — skips detection

[OisinKyne]

I'm unconvinced users want to be pushed to sell their (mostly insufficient) local inference on first install. I don't think there's pmf for it. If we want to make obol sell inference more able to auto detect different types of providers, that i could be down for, but i don't think this pr as is makes sense.

[bussyjd]

agreed — pivoted. removed the first-run wizard entirely (wizard.go deleted, -400 lines). the "earn first spend later" flow is gone.

kept the detection infra (detect.go — port scanning, server fingerprinting, model enumeration) and wired it into obol sell inference instead. now when you run obol sell inference myapi without --model, it scans local ports, shows what's running, and lets you pick. if --model is provided it skips detection entirely.

tldr: detection is opt-in at sell time, not pushed on first install.