fix(wp): anchor paged weights to pool_buf so view-only leaves get a backend

A paged weight consumed ONLY via a reshape/view (not a direct matmul src)
inherited buffer_id -1 from ggml_backend_sched_split_graph pass 4: the view's
id comes from its view_src (paged leaf, buffer==NULL -> -1), then the src loop
assigns the leaf = the view's still-(-1) id before the view's own fallback runs.
gallocr then asserts buffer_id>=0 on that leaf (before its data!=NULL skip). This
crashed every DeepSeek V4 paged load (wo_a = ggml_reshape_3d(...)->mul_mat), and
recurs for any arch that reshapes weights. Set paged weights' buffer=pool_buf at
init so the scheduler can name a backend; gallocr still skips them (data!=NULL),
eval_cb still re-patches ->data per op. Matches the steady state eval_cb already
creates (it sets ->buffer=pool_buf on every patch). Root-caused via codex+Fable
consults converging on ggml-backend.cpp:1217-1241.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014ZRfPpL8XFzk1hep9MMg9P

Author: kmbandy

src/llama.cpp

@@ -284,7 +284,17 @@ static bool init_weight_pager(llama_model & model, llama_model_loader & ml, cons
                 continue;
             }
             t->data   = placeholder;
-            t->buffer = nullptr;
+            // Anchor paged weights to the pool's device buffer so the backend
+            // scheduler can name a backend for them (buffer_id) even when a
+            // weight is reached only as a reshape/view's view_src — a leaf with
+            // buffer==NULL whose sole consumer is a view op falls through every
+            // split_graph assignment pass and gets buffer_id -1, tripping the
+            // gallocr assert (DeepSeek V4 wo_a = ggml_reshape_3d(...)->mul_mat).
+            // gallocr still skips it (is_allocated checks data != NULL first);
+            // eval_cb re-patches ->data (and already sets ->buffer = pool_buf)
+            // per op. This just brings load-time state to the post-first-decode
+            // steady state the eval_cb already produces.
+            t->buffer = pool_buf;
             ++n_placed;
         }
         LLAMA_LOG_INFO("%s: placeholder data set on %zu weight tensors (placeholder=%p, skipped %zu not in catalog)\n",