fix(unsloth): detect vision and thinking for the loaded model via /api/inference/status

[hwacookie]

Fixes #35, fixes #37

Problem

The OpenAI-compatible GET /v1/models exposes neither modality nor thinking metadata for Unsloth Studio:

• Vision models on Unsloth Studio are registered as text-only — images silently dropped #35 (vision): every model was registered as text-only, so Pi silently replaced attached images with (image omitted: model does not support images) — VLMs appeared to "have no vision", with no error or warning anywhere.
• Unsloth Studio: thinking-capable models are registered as non-reasoning — no thinking control in Pi #37 (thinking): the catalogue carries no reasoning metadata at all, so thinking-capable models were registered without reasoning.

Fix

Both capabilities are read from ONE local call that describes the loaded backend:

GET /api/inference/status  →  { active_model, is_vision, supports_reasoning, ... }

• active_model carries the same public id as the /v1/models entry (fallback: the loaded: true entry), so the flags attach to exactly one model.
• Vision (Vision models on Unsloth Studio are registered as text-only — images silently dropped #35): active model with is_vision: true → registered with input: ["text", "image"].
• Thinking (Unsloth Studio: thinking-capable models are registered as non-reasoning — no thinking control in Pi #37): active model with supports_reasoning: true → registered with reasoning: true.

Why not /api/models/check-vision/{id}? An earlier revision of this PR probed that endpoint for loaded models. For a model id that is not a resolvable HF repo (e.g. a local GGUF quant name like Qwen3.8-27B-IQ4_NL), Studio's handler falls back to fetching config.json from huggingface.co and 401s in a retry loop on EVERY listModels — and since Crossbar polls every 15s (Studio has no health endpoint), that became sustained HuggingFace traffic. /api/inference/status has no such fallback: local, instant, zero external requests.

Best-effort by contract: any probe failure (older Studio → 404, 401, refused, malformed body) degrades to the conservative pre-fix behaviour (text-only, no thinking), never throws, never blocks registration. Capabilities are known only for the loaded model; unloaded models pick up their real flags once they are loaded.

Verification

• Full suite: 711 passed / 8 skipped (pre-existing live-server skips); tsc --noEmit clean
• Regression tests assert the check-vision path is NEVER requested (trap routes included), so the HF-fallback regression cannot silently return
• is_vision on /api/inference/status verified against a running Studio instance (loaded Qwen3-VL reports is_vision: true)

Changes

• src/adapters/unsloth.ts — single /api/inference/status probe for vision + thinking; header docs
• tests/adapters/unsloth.fixture.ts — status route with is_vision; loaded-but-not-active entry
• tests/adapters/unsloth.test.ts — regression tests for both capabilities, incl. "never hits check-vision" guards
• CAPABILITY-MATRIX.md — Unsloth row updated