block_buffer: panic corrupts inline buffer position

Summary

### Summary

A caught panic may leave the cursor position of `EagerBuffer` or `ReadBuffer` in a corrupted state; this in turn allows out-of-bounds reads/writes.

### Details & PoC

The following two tests fail miri:

```rust #[cfg(miri)] #[test] fn eager_digest_blocks_panic_corrupts_inline_position() { // `EagerBuffer` stores its cursor in the last byte of the internal block. // When `digest_blocks` completes a previously partial block, it overwrites // that byte with input data before invoking the caller-provided `compress` // callback. If the callback panics, safe code can catch the panic and keep // using the buffer while its cursor byte no longer satisfies the internal // `pos < block_size` invariant. Under Miri this `get_pos` call reaches the // `unreachable_unchecked` used for the assumed-valid cursor. let mut buf = EagerBuffer::<U4>::new(&[1, 2]);

let _ = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| { buf.digest_blocks(&[3, 0xff], |_| panic!("simulated compression failure")); }));

let _ = buf.get_pos(); }

#[cfg(miri)] #[test] fn read_buffer_generator_panic_corrupts_inline_position() { // `ReadBuffer` stores its cursor in `buffer[0]`, but `write_block` gives // `gen_block` mutable access to the whole internal block before restoring // `buffer[0]` to a valid cursor. If `gen_block` writes an arbitrary first // byte and panics, safe code can catch the panic and later observe an // invalid cursor. Under Miri this `get_pos` call reaches the // `unreachable_unchecked` used for the assumed-valid cursor. let mut buf = ReadBuffer::<U4>::default();

let _ = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| { buf.write_block( 1, |block| { block[0] = 0xff; panic!("simulated block generation failure"); }, |_| {}, ); }));

let _ = buf.get_pos(); } ```

They fail on an `unreachable_unchecked!()` under the invariant for the `pos` to always be within bounds of the block.

### Impact

While the byte that overwrites `pos` may come from untrusted input and is therefore attacker-controlled, this still relies on the surrounding code catching the panic and carrying on, which should be uncommon in practice.

For this to be exploitable, the attacker also needs a way to trigger a panic here; I have not investigated how feasible that is.

### Credits

The issue was discovered by GPT-5.5