feat: Hebrew layerdrop replication spec + logit-path stitch wiring

src/gpu/distill_specs.yaml

@@ -527,3 +527,25 @@ heb-diac-small-s46:
   train: hebrew-v4/train.jsonl
   val: hebrew-v4/val.jsonl
   out: rababa_hebrew_distill_small/run-002-s46
+heb-diac-small-s46-layerdrop:
+  # Item 04 (TODO.publish-client): cross-lingual depth-cut test — the
+  # Arabic width/depth finding replicated on Hebrew. Single variable
+  # vs run-002-s46: encoder 12->6 via verbatim layer copy (layer_drop),
+  # same teacher, corpus, and epochs. Gate: the Arabic depth premium
+  # (1.21pp at 6ep) predicts a comparable lite-vs-full gap here.
+  teacher: rababa_hebrew/run-s46-phonikud-plus/run-002-gold-ft/best
+  teacher_volume: rababa
+  student_init: google/byt5-small
+  layer_drop: 'true'
+  student_config:
+    d_model: 1472
+    d_kv: 64
+    d_ff: 3584
+    num_heads: 6
+    enc_layers: 6
+    dec_layers: 4
+    feed_forward_proj: gated-gelu
+  train: hebrew-v4/train.jsonl
+  val: hebrew-v4/val.jsonl
+  out: rababa_hebrew_distill_small/run-003-s46-layerdrop
+  note: Hebrew lite rung; cross-lingual depth-cut replication

src/gpu/modal_distill.py

@@ -311,6 +311,10 @@ def distill(spec_id: str, epochs: int = 3, alpha: float = 0.5, temperature: floa
 
         cfg = T5Config(**spec["student_config"])
         student = T5ForConditionalGeneration(cfg).to(device)
+        if spec.get("layer_drop"):
+            # depth-cut students keep the pretrained init: verbatim
+            # layer copy, not the random init the tiny tier uses
+            _maybe_stitch(spec_id, spec, student)
     else:
         student = AutoModelForSeq2SeqLM.from_pretrained(spec["student_init"]).to(device)
     student.train()