Claude review: gpu: nova-core: mm: Add Virtual Memory Manager

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

Patch Review

This is the core mapping implementation with the prepare/execute split. The design is well thought out.

**Issue 1 (design question):** `MustUnmapGuard` uses `Cell<bool>`:
```rust
struct MustUnmapGuard {
    armed: Cell<bool>,
}
```
`Cell` is not `Sync`, which means `MappedRange` cannot be shared across threads. If this is intentional (mappings are always thread-local), document it. If not, consider `AtomicBool`.

**Issue 2:** In `ensure_pte_path`, the `and_then` closure:
```rust
self.pt_pages
    .get(&install_addr)
    .and_then(|p| Some(VramAddress::new(p.alloc.iter().next()?.offset())))
```
This could be simplified to `.map()` instead of `.and_then(|p| Some(...))`.

**Issue 3 (potential concern):** In `execute_map`, the PDE installation loop drains *all* prepared PT pages (`self.pt_pages`), not just those belonging to this specific mapping. If two `prepare_map` calls are interleaved before their `execute_map` calls, the first `execute_map` would install PDEs prepared for the second mapping. This seems intentional per the comment ("Shared by all pending maps in the Vmm"), but it means `Vmm` effectively serializes all map operations. Worth documenting this constraint explicitly.

**Issue 4:** The `map_pages` convenience wrapper validates `va_range` against `pfns.len() * PAGE_SIZE`, which is good. The `checked_mul` + `into_safe_cast` chain handles overflow correctly.

---
Generated by Claude Code Patch Reviewer