feat: multi-song architecture + top-7 preset re-bake + tar/gra EQUIV

feat: data-driven Mandarin/long-vowel EQUIV additions for Castilian

feat: w1-doubled geom, drop n from m-EQUIV, incremental render+score loop

feat: eval metric — w2v cache, geom F1, RAW prec with short-window floor, phrase_coverage multiplier

feat: data-driven phoneme alts + non-greedy match + re-bake to new tops

feat: language-aware Castilian metric + buenas_tardes preset + bake validator

feat: ship 3 best-single bakes (consistency over composite)

fix: tighten ɝ EQUIV + re-bake happy_birthday from composite

feat: Castilian-aware eval metric + re-bake buenos_dias and happy_birthday

feat: re-bake hap-pee birth-day with composite_lev_xf

chore: bump SoulX-Singer submodule to 49a9c27 (soundfile audio_utils)

fix: install ffmpeg system-level for torchcodec compatibility

fix: bump torch to 2.11.0 for ZeroGPU compatibility

feat: re-bake hap-pee birth-day as crossfaded composite

feat: fix first "happy" in hap-pee birth-day preset

docs: rewrite README for the shipped AIchael Jackson demo

feat: thread --seed through render() for reproducible audio + bake seed=100 buenos dias preset

feat: bump default melisma threshold to 0.30 + bake matching presets

feat: melisma_mode advanced setting (off | default)

feat: default n_steps to 16 + bake matching buenos-dias cache

feat: swap presets to hap-pee birth-day and syn-thet tic-voice

fix: cache check before @GPU so cache hits don't burn ZeroGPU quota

feat: bake buenos-dias preset into assets/cache for instant first-load

perf: 30s ZeroGPU cap + dedup cache for repeat syllable renders

fix: drop ZeroGPU duration to 60s to stay under per-call cap

perf: use cuda + fp16 when available (HF ZeroGPU was running on cpu)

fix: bind 0.0.0.0 + disable SSR for HF Spaces healthcheck

fix: pin python_version to 3.11 for HF Spaces (torch==2.2.0 has no 3.13 wheels)

fix: drop sdk_version so HF Spaces picks supported gradio default

fix: trim short_description to 60-char HF Spaces limit

HF Spaces scaffolding: submodule, bootstrap, frontmatter

Milestone 15: AIchael Jackson — first Gradio demo

Milestones 13/14: double-B plosive recipe + re-mixed cover

Milestone 12: buenos dias cover using milestone 11 vocal

Milestone 11: perturbed chorus prompt avoids leakage at lower cost

Milestone 10: buenos dias vocal + Thriller accompaniment cover

Milestone 9: validate voice cloning with buenos dias target

Milestone 8: MJ timbre directly from SoulX prompt, no Seed-VC needed

Initial commit + Milestone 7: MJ-derived synth from real Thriller chorus