Claude review: gpu/buddy: Track per-order free blocks with a scoreboard

[Claude Code Review Bot <claude-review@example.com>]

Patch Review

**Correctness: The scoreboard maintenance looks correct.**

The increment/decrement sites are:

1. `mark_free()` — increment (block enters free state)
2. `mark_allocated()` — decrement (block leaves free state)
3. `mark_split()` — decrement (block leaves free state)
4. `__gpu_buddy_free()` coalescing loop — decrement for buddy being removed from tree
5. `__force_merge()` — decrement for block being removed from tree
6. `__gpu_buddy_undo_splits()` — decrement for block being removed from tree

These cover all the places where `rbtree_remove` is called on a free block (decrements) and where `rbtree_insert` is called on a free block (increments, via `mark_free`).

**Initialization and cleanup look correct** — `kcalloc` zeroes the array, `kfree` in both the init error path and `fini`.

**One minor note on the print threshold change:**

```c
-		if (free < SZ_1M)
+		if (free < SZ_1M)
```

In this patch the threshold is still based on `free` (the total free bytes). This gets changed to `block_size` in patch 5. That's fine — incremental development.

No issues.

---
Generated by Claude Code Patch Reviewer