AI v2 C2: BiRefNet matting upgrade for image-to-3D (Fast/Best tiers) #1016

Part of #818 (Track C2). One of the epic's three named quick wins. Status: ABSENT.

BackgroundRemover is still U²-Net-only at 320² (rembg/u2net.onnx). BiRefNet is MIT (code + weights), gives 1024² mattes with far better edges, and has ready ONNX exports (~930 MB fp32 / ~230 MB quantized; -lite variants exist).

Scope

Drop-in behind the same BackgroundRemover interface: add a Quality::{Fast,Best} tier (the TripoSR tier precedent), rembg/birefnet.onnx, keep U²-Net as the small/fallback tier. New row in THIRD_PARTY_AI_MODELS.md in the same PR.

Acceptance

Side-by-side on 5 test photos: no halo ring in the reconstruction on at least the hair/fur case that fails today.

Note on scope (measured 2026-09)

A related-sounding matte bug was investigated and turned out not to affect the app: the shipped U²-Net path produces clean mattes (13.9% / 16.1% coverage, 0 interior-hole pixels) on the images that were generating bad faces — see the correction on #998. The threshold-matte failure is scoped to standalone trellis-cli. So this issue is a genuine quality upgrade (edge fidelity at higher resolution), not a bug fix, and should be scheduled on those terms.