Claude review: rust: auxiliary: generalize Registration over ForLt

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

Patch Review

Converts `Registration<T: 'static>` to `Registration<'a, F: ForLt + 'static>`. Two constructors:

1. `unsafe fn new_with_lt()` — for borrowed data, caller must not `mem::forget()` the Registration
2. `fn new()` — safe, requires `F::Of<'a>: 'static`

The `PhantomData<(fn(&'a ()) -> &'a (), F)>` uses the `fn(&'a ()) -> &'a ()` trick to make `'a` invariant (not covariant or contravariant), which is correct for a type that holds data borrowed for `'a`.

The `registration_data()` method now uses `TypeId::of::<F>()` (the ForLt type itself, not `F::Of<'static>`) for the type check. This is correct because `F` is `'static` (trait bound) and uniquely identifies the type family.

The transmute in `new_with_lt`:
```rust
let data: Pin<KBox<RegistrationData<F::Of<'static>>>> =
    unsafe { core::mem::transmute(data) };
```
This widens `F::Of<'a>` to `F::Of<'static>` for storage. The safety argument is that `Registration`'s Drop will release this data before `'a` ends (enforced by the unsafe contract on `new_with_lt`). Combined with ForLt's covariance, the `cast_ref` in `registration_data()` can safely shorten back from `'static` to the borrow scope.

**The `mem::forget` soundness hole is well-addressed** by splitting into safe/unsafe constructors. The safe `new()` has `F::Of<'a>: 'static`, meaning forgetting is harmless. The unsafe `new_with_lt()` documents the requirement.

The `Send`/`Sync` bounds use `for<'a> F::Of<'a>: Send` which is correct — the data must be Send for all possible lifetimes since the actual stored lifetime is erased to `'static`.

---

---
Generated by Claude Code Patch Reviewer