fim: apply fuzzy backtrack offset to all completions

Store the original (untrimmed) cached responses and the fuzzy backtrack
offset. At render time, trim ALL completions by the same offset so that
cycling through alternatives shows correctly trimmed suggestions instead
of the full untrimmed content.

Assisted-by: llama.cpp:local pi

Author: ggerganov

autoload/llama.vim

@@ -1009,6 +1009,8 @@ function! s:fim_try_hint(pos_x, pos_y)
     " Check if the completion is cached (and update LRU order)
     let l:raw = s:cache_get(l:hash)
     let l:selected = 0
+    let l:fuzzy_offset = -1
+    let l:fuzzy_typed = ''
 
     " ... or if there is a cached completion nearby (128 characters behind)
     " Looks at the previous 128 characters to see if a completion is cached. If one is found at (x,y)
@@ -1047,14 +1049,12 @@ function! s:fim_try_hint(pos_x, pos_y)
                         continue
                     endif
 
-                    " build a modified responses list with the matched completion trimmed
-                    let l:modified = copy(l:responses)
-                    let l:modified[c_idx] = extend(copy(l:responses[c_idx]), {'content': l:remainder})
-
                     if len(l:remainder) > l:best_len
                         let l:best_len = len(l:remainder)
-                        let l:best_raw = json_encode(l:modified)
+                        let l:best_raw = l:response_cached  " store original (untrimmed)
                         let l:best_selected = c_idx
+                        let l:fuzzy_offset = i
+                        let l:fuzzy_typed = l:removed
                     endif
                 endfor
             endif
@@ -1073,11 +1073,20 @@ function! s:fim_try_hint(pos_x, pos_y)
             let l:selected = 0
         endif
 
+        " apply fuzzy offset: trim first (fuzzy_offset + 1) chars from each completion
         let l:completions = []
         for l:resp in l:responses
-            call add(l:completions, split(get(l:resp, 'content', ''), "\n", 1))
+            let l:content = get(l:resp, 'content', '')
+            if l:fuzzy_offset >= 0 && len(l:content) > l:fuzzy_offset
+                let l:content = l:content[l:fuzzy_offset + 1:]
+            endif
+            call add(l:completions, split(l:content, "\n", 1))
         endfor
 
+        " store fuzzy info so cycling applies the same trimming
+        let s:fim_data['fuzzy_offset'] = l:fuzzy_offset
+        let s:fim_data['fuzzy_typed'] = l:fuzzy_typed
+
         call s:fim_render(l:pos_x, l:pos_y, l:completions, l:selected, l:responses)
 
         " run async speculative FIM in the background for this position