v3.46-trained: AMS_TRAINED_WEIGHTS shape-mismatch is warn+skip, not fatal

Root cause: 4.25 prefix_length_scaling_probe intentionally builds model_b with
L_mem doubled (default 8 -> 16).  The checkpoint was trained with L_mem=8, so
L_mem-dependent tensors (e.g. mem_tokens[L_mem, d_LLM]) legitimately don't fit
model_b — this is not a corrupt/incompatible ckpt, it's a deliberate Cfg scan.

Old behavior: raise RuntimeError on any shape mismatch -> errored 4.25.
New behavior:
  - Per-tensor shape mismatch is logged and skipped (first 5 detailed, rest summarized).
  - Hard failure only when the ckpt had non-backbone content (>10 tensors) AND
    zero tensors loaded — that is the §6 'wrong-SUT ckpt' pattern we must catch.

Keeps the §6 protection against loading v344_trained.pt / v348_stacked.pt against
a v3.46 SUT (they would mostly shape-mismatch and hit the loaded==0 guard), while
letting L_mem-scaling probes proceed.

Co-authored-by: FluffyAIcode <FluffyAIcode@users.noreply.github.com>

Author: cursoragent

scheme_b_v344.py

@@ -2541,14 +2541,23 @@ def _maybe_load_trained_weights(self):
                 else:
                     skipped += 1
         prov = blob.get("provenance", "?") if isinstance(blob, dict) else "?"
+        total_nonbb_ckpt = sum(1 for k in sd if not k.startswith("backbone"))
         print(f"  [AMS_TRAINED_WEIGHTS] loaded={loaded} skipped={skipped} "
               f"shape_errs={len(shape_errs)}  path={path}  provenance={prov}")
         if shape_errs:
             for n, s_model, s_ckpt in shape_errs[:5]:
-                print(f"    ! shape mismatch {n}: model={s_model} ckpt={s_ckpt}")
+                print(f"    ! shape mismatch (skipped) {n}: model={s_model} ckpt={s_ckpt}")
+            if len(shape_errs) > 5:
+                print(f"    ... and {len(shape_errs) - 5} more shape mismatches, all skipped")
+        # Raise only if essentially nothing loaded AND the ckpt had content to offer:
+        # this catches the "loaded a v344/v348 ckpt against v3.46 shapes" mistake
+        # warned about in SPRINT_CLOSEOUT_v3.46.md \u00a76, without breaking probes
+        # like 4.25 that scale L_mem and legitimately have a few mismatching tensors.
+        if loaded == 0 and total_nonbb_ckpt > 10:
             raise RuntimeError(
-                f"AMS_TRAINED_WEIGHTS shape mismatch on {len(shape_errs)} tensor(s); "
-                f"ckpt not compatible with current SUT shapes")
+                f"AMS_TRAINED_WEIGHTS loaded 0 non-backbone tensors "
+                f"(ckpt had {total_nonbb_ckpt}); shape_errs={len(shape_errs)}. "
+                f"ckpt appears incompatible with current SUT shapes")
 
     def _compute_filler_centroid(self):
         if self.content_classifier is None or self.backbone is None: