Release 0.2.0

[TroyHernandez]

CRAN release prep. Check record on the submission tarball:

• Linux (as-cran): 0 errors, 0 warnings, 2 NOTEs (new submission; torch load-hook noise in check children — local build artifact, absent on Windows)
• Windows R 4.6.0 (as-cran, full test suite executed): 0 errors, 0 warnings, 1 NOTE (new submission)

In this release branch

• serve(): zero-dependency HTTP server (base R sockets, whisper/chatterbox skeleton, port 7812) — /v1/images/generations (flux2/zimage/flux1), /v1/videos/generations (ltx), /health. Hardened: bearer-token auth, pixel/frame limits, bounded LRU of connector embeds (never raw Gemma stacks), CUDA-OOM exits for clean supervisor restart, never downloads. Smoke-tested live twice (happy path + auth/caps path); 20 pure-R unit tests for parser/router/handlers. systemd unit in inst/diffuser.service.
• Consent everywhere: legacy download_model()/download_component() now route through the same consent gate as the modern download functions; README claim corrected to match.
• SD tiers relabeled fp16 (no SD quantized weights exist; placement varies, not precision) with tests updated.
• Performance Levers vignette (three axes: precision ladder, placement, residency) + README hardware-requirements table; accuracy pass on recommend() consumption claims.
• NEWS consolidated into one 0.2.0 section; version 0.2.0.
• Doc/check fixes: roxygen pairing, return_penultimate, quantizer + constructor NULL-defaults (Rd width clean), test fixtures excluded from the tarball (25 fixture suites skip gracefully when absent; full parity runs locally/CI).