[ronaldtse]

TODO.sota substrate migration: all training/labeling/export scripts now
run on HF Jobs (datasets + buckets), Modal retired for compute (used
only to export existing corpora/checkpoints one time).

• train_hebrew_plane_hf.py — run-022 byt5-base scale-up (gate cleared:
  8.31 nakdimon DER in-job vs 12.48 shipped)
• export_plane_onnx_hf.py — plane.onnx fp32+int8 export for run-022
• label_arwiki_hf.py — r7 teacher pseudo-labels arwiki with margin
  filter (WO06 stage 1; NO LLM labels per standing rule)
• train_arabic_r8_hf.py — r8 noisy-student, gold+pseudo mixture,
  in-job SadeedDiac-25 gate (stage 2)
• train_thai_tiny_hf.py — WO08 end-to-end: wiki slice, umt5 teacher
  KD, ~12M student, PER+size gates
• probe_hebrew_audio_pseudolabel.py — WO09 parked probe (ASR stage)
• scripts/build_hebrew_v4_local.py — local v4 corpus rebuild