forest-n-leaf is not updated by pruning!, so it reports the pre-pruning leaf count

[masatoi]

The forest-n-leaf slot is written only by make-forest and make-regression-forest. pruning! calls set-leaf-index-forest!, which updates each dtree-max-leaf-index, but nothing updates forest-n-leaf. The exported accessor therefore keeps reporting the pre-pruning leaf count.

Since forest-n-leaf is the natural accessor to reach for when measuring how much Global Pruning compacted the model — one of the features advertised in the README — this silently reports zero reduction.

Reproduction

SBCL 2.6.7, master @ b4e5ec5. Same setup as #14 (30 trees, 2000 synthetic data, pruning-rate 0.2):

(clrf:forest-n-leaf *forest*)  ; => 2759

;; ground truth
(reduce #'+ (mapcar #'cl-random-forest/src/random-forest::dtree-max-leaf-index
                    (clrf:forest-dtree-list *forest*)))  ; => 2759

(clrf:pruning! *forest* *learner* 0.2)

(clrf:forest-n-leaf *forest*)  ; => 2759   <- stale, unchanged

(reduce #'+ (mapcar #'cl-random-forest/src/random-forest::dtree-max-leaf-index
                    (clrf:forest-dtree-list *forest*)))  ; => 2542   <- actually pruned

Counting the leaves by traversal with do-leaf agrees with dtree-max-leaf-index (2542), so pruning itself is working correctly — only the cached slot is wrong.

Cause

make-forest (src/random-forest.lisp:660-662) is the only writer:

(setf (forest-n-leaf forest)
      (apply #'+ (mapcar #'dtree-max-leaf-index
                         (forest-dtree-list forest))))

pruning! (src/random-forest.lisp:1219-1226) ends with (set-leaf-index-forest! forest) and does not touch forest-n-leaf.

Impact

Cosmetic / reporting only, as far as I can tell. make-refine-learner (src/random-forest.lisp:816) and make-regression-refine-learner (src/random-forest.lisp:1053) both compute input-dim by summing dtree-max-leaf-index directly rather than reading forest-n-leaf, so the refine pipeline is unaffected. grep shows forest-n-leaf has no other readers inside src/.

Suggested fix

Move the sum into set-leaf-index-forest! so every caller (both make-forest and pruning!) keeps it consistent:

(defun set-leaf-index-forest! (forest)
  (dolist (dtree (forest-dtree-list forest))
    (set-leaf-index! dtree))
  (let ((sum 0)
        (offset (forest-index-offset forest)))
    (setf (aref offset 0) 0)
    (loop for dtree in (forest-dtree-list forest)
          for i from 1 below (forest-n-tree forest)
          do (setf sum (+ sum (dtree-max-leaf-index dtree)))
             (setf (aref offset i) sum)))
  (setf (forest-n-leaf forest)
        (reduce #'+ (mapcar #'dtree-max-leaf-index (forest-dtree-list forest)))))

and drop the now-redundant setf in make-forest / make-regression-forest.