docs: document 16k window default + indel baseline caveat

docs: note 16k window + indel-score caveat in app Documentation tab

perf: default CONTEXT_LEN to 16k (~2.5x faster CPU inference; NTV3_CONTEXT_LEN overrides)

feat: make CONTEXT_LEN env-configurable via NTV3_CONTEXT_LEN (default 32k)

perf(app): full-output cache — cached example clicks return in ~50ms (vs ~10s)

fix(app): pre-warm example cache in background thread (was hitting startup timeout)

perf(app): pre-warm lru_cache for built-in examples at startup

perf(app): lru_cache on per-variant inference — example clicks now instant after first run

deps: add py2bit for hg38.2bit support

feat(spaces): fast hg38.2bit local lookup with auto-download from HF dataset

fix(spaces): use Ensembl REST as fallback for human (HF egress blocks UCSC)

ux: show explicit error when sequence fetch fails

debug: log UCSC/Ensembl fetch failures + surface them in UI

fix(spaces): defensive monkey-patch for gradio_client schema bool bug (No API found)

fix(spaces): bump gradio sdk_version to 6.13.0 + correct suggested_hardware key

Pin huggingface_hub<1.0 (HfFolder removed in 1.0)

Add audioop-lts for Python 3.13 (pydub/gradio compat)

Initial deploy: MAGI variant interpreter (gradio_app)

initial commit