fix: address codex review round 2

- spill-policy reserves the spill notice's byte cost inside maxInlineBytes, so
  the replacement (preview + notice) never exceeds the documented model-facing
  cap. When the notice alone fills the budget the preview is empty; when even a
  notice-only replacement is not smaller than the original, the inline result is
  kept (spilling would only add bytes).
- retention TextRetainer trims an oversized single suffix chunk to the last
  suffixCap bytes on push, so tail/headTail retention stays bounded by suffixCap
  instead of retaining and re-copying the whole chunk in finish() — this is the
  spill preview path, which pushes the whole result as one chunk.

packages/util/retention/src/index.ts

@@ -355,6 +355,19 @@ export class TextRetainer {
         this.suffixHeld -= head.length
         head = this.suffixChunks[0]
       }
+      // The head chunk can still hold leading bytes beyond the last `suffixCap`
+      // — a single chunk LARGER than the window is retained whole by the loop
+      // above (dropping the only chunk would leave < cap). Trim those leading
+      // bytes so the accumulator (and finish()'s concat) stays bounded by
+      // `suffixCap` instead of allocating/copying the full chunk again;
+      // finish() only ever reads the last `suffixLen ≤ suffixCap` bytes, so this
+      // drops nothing it would return. (head.length > excess by the loop
+      // invariant `suffixHeld - head.length < suffixCap`, so the slice is non-empty.)
+      if (head !== undefined && this.suffixHeld > this.suffixCap) {
+        const excess = this.suffixHeld - this.suffixCap
+        this.suffixChunks[0] = head.subarray(excess)
+        this.suffixHeld -= excess
+      }
     }
 
     // Dropped = bytes that no side can keep. Compute cumulative omission the