fix: corda silently ran as plain SVD; wire calibration + persist data-driven residual

The benchmark only passed calibration_data to eva, so antipasto_corda's
group_init hit `if calibration_data is None: return` and every corda run was
actually plain SVD. The covariance orientation never executed -- all prior
corda-vs-antipasto comparisons are void.

- antipasto_corda.group_init: raise on None instead of silently degrading
  (orientation is the variant's whole identity; fail loud).
- benchmark: feed ~256 MetaMath calibration samples (IPM, per PEFT/CorDA) to
  corda and to cov_orient ablate; run_id now carries an __lr tag.
- adapter.save/load: a data-driven group_init rewrites the frozen base residual
  W_res into a form init() cannot reproduce at load (it only knows the plain
  top-r crop). Persist those residuals in the adapter and restore them. Fixes a
  reload-logits mismatch that was masked while group_init never ran.
- probe check: compare every saved tensor (lora_ buffers AND base residuals)
  against the reloaded model state.
- justfile: bench-variant gains an lr_override (the core wants a tamer lr than
  the gain's 5e-3).

Co-Authored-By: Claudypoo <noreply@anthropic.com>

src/lora_lite/adapter.py

@@ -63,6 +63,7 @@ def attach(model: nn.Module, cfg: AdapterConfig, calibration_data=None, *, _skip
         attached_targets.append((name, layer, role))
 
     group_init = getattr(variant, "group_init", None)
+    ran_data_init = group_init is not None and not _skip_group_init and calibration_data is not None
     if group_init is not None and not _skip_group_init:
         group_init(model, attached_targets, cfg, calibration_data)
 
@@ -72,7 +73,13 @@ def attach(model: nn.Module, cfg: AdapterConfig, calibration_data=None, *, _skip
         else:
             handles.append(layer.register_forward_hook(_hook))
 
-    setattr(model, _ATTACHED_ATTR, {"cfg": cfg, "targets": attached_names, "handles": handles})
+    # A data-driven group_init (CorDA orient, Wanda re-select) rewrites the frozen
+    # base residual W_res into a form init() cannot reproduce at load time (it only
+    # knows the plain top-r crop). So those residuals are part of the saved adapter.
+    base_weight_keys = [f"{n}.weight" for n in attached_names] if ran_data_init else []
+    setattr(model, _ATTACHED_ATTR,
+            {"cfg": cfg, "targets": attached_names, "handles": handles,
+             "base_weight_keys": base_weight_keys})
     return handles
 
 
@@ -102,7 +109,11 @@ def save(model: nn.Module, path: str) -> None:
     state = getattr(model, _ATTACHED_ATTR, None)
     if state is None:
         raise RuntimeError("no adapter attached; call attach() first")
-    sd = {k: v.detach().cpu() for k, v in model.state_dict().items() if "lora_" in k}
+    full_sd = model.state_dict()
+    sd = {k: v.detach().cpu() for k, v in full_sd.items() if "lora_" in k}
+    # data-driven variants also persist their rewritten base residuals (see attach()).
+    for wk in state.get("base_weight_keys", []):
+        sd[wk] = full_sd[wk].detach().cpu()
     metadata = {"cfg": json.dumps(state["cfg"].to_dict())}
     from safetensors.torch import save_file
     save_file(sd, path, metadata=metadata)

src/lora_lite/variants/antipasto_corda.py

@@ -92,14 +92,18 @@ class AntiPaSTOCorDA:
     def group_init(model: nn.Module, targets, cfg, calibration_data: CalibrationData | None) -> None:
         """Re-orient each target's SVD by its input covariance C = E[x x^T].
 
-        Without calibration_data the plain-SVD init from init() is kept (so this
-        degrades to antipasto, rotation-free).
+        Covariance orientation IS this variant's identity, so calibration_data is
+        mandatory -- fail loud rather than silently degrade to plain SVD (which is
+        just antipasto and was the bug that made every corda run a no-op).
 
         Called by attach() BEFORE any training, so the trainable g is still at its
         zero init when the basis changes -- re-orienting zero gains is a no-op, no
         re-indexing needed. Do not call group_init after training has updated g."""
         if calibration_data is None:
-            return
+            raise ValueError(
+                "AntiPaSTOCorDA requires calibration_data (covariance orientation is "
+                "its whole point); got None. Pass attach(model, cfg, calibration_data=...)."
+            )
 
         layers = {name: layer for name, layer, _ in targets}
         # accumulate C = sum x x^T on CPU. Peak GPU cost would otherwise be