memory: snapshot — R49 1-epoch pilot complete and awaiting adjudication; MeroMero A4B quantized, v2 blocked; althing 3.6.2 rolled

Ships the two pending code changes alongside the memory that describes them:
train_voice_lora.py gains --eval-steps/--save-steps (the 3-epoch pilot overfit
with per-epoch eval and save_strategy="no", so the minimum was neither visible
nor recoverable), and the althing post-office compose is pinned to 3.6.2.

Index rewritten: in-flight compressed from 176 lines to the live state, three
detail files added for R49 D1-D3, the MeroMero four-failure chain, and the
althing rollout. Seven closed pre-08-27 entries archived; the guards held back
the rest because they carry open deferred pointers.

scripts/r49-corpus/train_voice_lora.py

@@ -80,6 +80,14 @@ def main() -> int:
     ap.add_argument("--batch", type=int, default=1)
     ap.add_argument("--accum", type=int, default=8)
     ap.add_argument("--seed", type=int, default=4919)
+    # Step-wise eval + save. The 3-epoch pilot showed held-out loss rising every
+    # epoch (3.198 -> 3.318 -> 3.385) while train loss fell, so the optimum is
+    # EARLIER than one epoch-boundary eval can see, and `save_strategy="no"` left
+    # nothing to fall back to. Evaluate and checkpoint on a step grid so the
+    # minimum is located from data instead of guessed, and so the best adapter
+    # actually exists on disk when it is found.
+    ap.add_argument("--eval-steps", type=int, default=0, help="0 = per-epoch")
+    ap.add_argument("--save-steps", type=int, default=0, help="0 = no intermediate saves")
     a = ap.parse_args()
 
     torch.manual_seed(a.seed); random.seed(a.seed)
@@ -132,6 +140,7 @@ def main() -> int:
     prov = {"run": "r49-h02-pilot", "base": a.base, "corpus": a.corpus, "corpus_sha256_16": corpus_sha,
             "seq_len": a.seq_len, "lora_rank": a.rank, "lora_alpha": 2 * a.rank, "targets": TARGETS,
             "lr": a.lr, "epochs": a.epochs, "batch": a.batch, "grad_accum": a.accum, "seed": a.seed,
+            "eval_steps": a.eval_steps or "per-epoch", "save_steps": a.save_steps or "none",
             "train_blocks": len(train_blocks), "train_tokens": tr_tok, "val_blocks": len(val_blocks),
             "trainable_params": trainable, "total_params": total,
             "trainable_pct": round(100 * trainable / total, 3),
@@ -151,7 +160,12 @@ def main() -> int:
         # longer exists. Read the signature, do not assume the 4.x one.
         lr_scheduler_type="cosine", warmup_steps=max(1, int(0.03 * steps_per_epoch * int(a.epochs))),
         bf16=True, logging_steps=10,
-        save_strategy="no", eval_strategy="epoch", report_to=[], seed=a.seed,
+        save_strategy=("steps" if a.save_steps else "no"),
+        save_steps=(a.save_steps or 500),
+        save_total_limit=12,
+        eval_strategy=("steps" if a.eval_steps else "epoch"),
+        eval_steps=(a.eval_steps or None),
+        report_to=[], seed=a.seed,
         gradient_checkpointing=True, dataloader_num_workers=2,
     )
     trainer = Trainer(model=model, args=args, train_dataset=Packed(train_blocks),