perf(lora): oversample the hosted-factor sketch to near-exact svd

Sketch hosted deltas at rank+64 with eight power iterations and keep the
top rank columns; this lands within noise of the exact decomposition at
roughly twice a sketch cost the factor cache pays once per configuration.
Bump the cache format so narrower-sketch entries reload as misses.

modules/lora/lora_factor_cache.py

@@ -96,7 +96,7 @@ def begin_pass(wanted_names):
             from safetensors import safe_open
             with safe_open(path, framework='pt', device='cpu') as f:
                 meta = f.metadata() or {}
-                if meta.get('sig') == sig and meta.get('fmt') == '2':
+                if meta.get('sig') == sig and meta.get('fmt') == '3':
                     for k in f.keys():
                         entries[k] = f.get_tensor(k)
             os.utime(path, None) # freshness for LRU eviction
@@ -190,7 +190,7 @@ def flush():
         from safetensors.torch import save_file
         os.makedirs(cache_root, exist_ok=True)
         tmp = state['path'] + '.tmp'
-        save_file(state['store'], tmp, metadata={'sig': state['sig'], 'fmt': '2'})
+        save_file(state['store'], tmp, metadata={'sig': state['sig'], 'fmt': '3'})
         os.replace(tmp, state['path'])
         evict()
     except Exception as e:

modules/lora/lora_sdnq.py

@@ -274,7 +274,9 @@ def apply_hosted(self, network_layer_name, updown, wanted_names, use_previous=Fa
     # svd_lowrank draws random projections; fork so user generation seeds are untouched and re-applies are deterministic
     with torch.random.fork_rng(devices=[D.device] if D.device.type == 'cuda' else []):
         torch.manual_seed(0)
-        U, S, V = torch.svd_lowrank(D, q=q, niter=4)
+        # oversampled sketch with extra power iterations lands within noise of exact svd; only the top q columns are kept
+        U, S, V = torch.svd_lowrank(D, q=min(q + 64, *D.shape), niter=8)
+    U, S, V = U[:, :q], S[:q], V[:, :q]
     energy = float(S.square().sum() / D.square().sum().clamp(min=1e-30)) # captured fraction, in the weighted domain when calibrated
     up_h = (U * S).to(dtype=dtype)
     down_h = V.t()