llama.cpp patches out-of-bounds read in LoRA loading

llama.cpp shipped build b10451 on August 16, closing an out-of-bounds read in LoRA adapter loading. PR #27056 adds the fix. A bounds check now requires LoRA tensor data to sit inside the adapter file before the loader reads it, per the release notes. It's one check on one loading path, not a rewrite of the loader.

LoRA adapters get downloaded from model hubs constantly. Apps that embed llama.cpp for local inference load those files as inert data. Any app that let a user point at a third-party adapter was trusting a file its loader never checked before b10451.

The release notes don't say whether the fix gets backported to older release branches. Nor do they say whether triggering the read needs a malicious adapter or just a malformed one.

The fix landed only in build b10451, posted August 16 on GitHub. Teams running an older pin must update it before loading another third-party adapter.