feat(distill): G2a epochs-rung + gentle d768 stitch rung

src/gpu/distill_specs.yaml

@@ -164,6 +164,58 @@ ara-diac-small-layerdrop:
     verbatim (both width-cut rungs failed); int4 target ~75MB; Muon
     optimizer per E3
   optimizer: muon
+ara-diac-small-2-6ep:
+  # G2a: the residual-attribution disambiguation - same as the 2.0 rung
+  # (r7 labels, Muon, vanilla ByT5-small) at 6 epochs instead of 3.
+  # If no gain, the 2.25pp residual is NOT undertraining -> domain.
+  teacher: rababa_arabic_byt5/run-007-news/best
+  teacher_volume: rababa
+  out_volume: rababa
+  student_init: google/byt5-small
+  train: r5-units/domain.txt
+  train_extra:
+  - r5-units/replay.txt
+  unit_limits:
+  - 24000
+  - 6000
+  max_len: 1450
+  label_beams: '1'
+  out: rababa_arabic_distill_small/run-007-r7-muon-6ep
+  labels_file: teacher_labels_r7.jsonl
+  labels_complete: 'true'
+  mode: sequence
+  optimizer: muon
+  note: G2a epochs-rung; gate vs 4.8218 - flat means domain-limited
+ara-diac-small-d768:
+  # G2b companion + frontier point: the UNTESTED gentle width cut
+  # (2x, not 3.8x) with SVD stitch + Muon. ~105M params, int8 ~105MB.
+  # Closes the "you never tried a gentle stitch" reviewer attack.
+  teacher: rababa_arabic_byt5/run-007-news/best
+  teacher_volume: rababa
+  out_volume: rababa
+  student_init: google/byt5-small
+  student_config:
+    d_model: 768
+    d_kv: 64
+    d_ff: 1856
+    num_heads: 12
+    enc_layers: 12
+    dec_layers: 4
+    feed_forward_proj: gated-gelu
+  train: r5-units/domain.txt
+  train_extra:
+  - r5-units/replay.txt
+  unit_limits:
+  - 24000
+  - 6000
+  max_len: 1450
+  label_beams: '1'
+  out: rababa_arabic_distill_small/run-008-d768-stitch
+  labels_file: teacher_labels_r7.jsonl
+  labels_complete: 'true'
+  mode: sequence
+  optimizer: muon
+  note: gentle-stitch rung (2x cut); frontier point at ~105M
 ara-diac-tiny-stitched:
   teacher: rababa_arabic_byt5/run-006-morph/best
   teacher_volume: rababa