gemma4: full feature mirror resync after prefix cache restore

After restoring KV from a snapshot, do_prefill only syncs the feature mirror
for the delta tokens [snap_pos..committed). The positions [0..snap_pos) in
the mirror retain stale data from the previous request's decode phase (which
may have diverged from the current prompt context after the ring buffer
wraps).

Fix: call draft_feature_mirror_sync_tail after restore to resync the entire
[0..committed) feature range from cache_.target_feat to the mirror. This
ensures the draft model sees consistent features and maintains high
acceptance rate (AL) during speculative decoding.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: howard0su

dflash/src/gemma4/gemma4_backend.cpp

@@ -671,6 +671,14 @@ GenerateResult Gemma4Backend::restore_and_generate(int slot,
         cache_.cur_pos = committed;
     }
 
+    // Full feature mirror resync after restore: do_prefill only synced the
+    // delta [snap_pos..committed). Re-sync the entire [0..committed) range so
+    // the draft model sees correct features for the full context.
+    if (feature_mirror_.target_feat && cache_.target_feat && !draft_parked_ && committed > 0) {
+        draft_feature_mirror_sync_tail(cache_.target_feat, cache_.target_feat_cap,
+                                       feature_mirror_, committed);
+    }
+
     // Generate
     if (req.n_gen > 0) {
         const bool can_spec = dflash_target_