perf(hf-space): pre-load model at module init (Option 3 refactor)

fix(parser): bump quoted_span limit 200 → 400 chars

fix: address parallel code-review findings (security + quality + spec)

fix(security): pass user Anthropic key directly to SDK, never via env

test(hf-space): cover _call_huggingface (full) + _call_zerogpu (refactor)

test(hf-space): cover diagnose() Premium path + Anthropic call shape

feat(hf-space): add ZeroGPU backend for HuggingFace Pro Spaces

feat(hf-space): add HuggingFace backend (Gemma/Phi-4) alongside Anthropic

feat(hf-space): Gradio diagnostic with Anthropic SDK + prompt caching (US4)