Matt Caswell

This looks good to me. I'll hold back from "approving" until #19703 is merged and this is subsequently rebased.

Hmmm....what changed to break this?

I was wondering about whether we could do "normal" handshakes, but provide fuzzed data into the fault injector somehow.

FYI, I'm currently trying to look at this. I'm having problems though when building for fuzzing. The fuzzing README says you should configure like this: ```` CC=clang ./config enable-fuzz-libfuzzer \...

I've made some significant progress on getting the code for the quic-client fuzzer to actually work, but I'm still encountering linker issues building it to actually do some fuzzing. My...

> It's a bit unclear to me how you can get in a situation where that's not available. On Debian, it's currently part of libstdc++-13-dev, where -13 is the gcc...

See #22368. I've taken the commit from this PR, made a few tweaks, and added additional commits to enable the fuzzer to successfully complete the handshake/read data.

> But I don't think that is the reason. It is the reason. The real .h files don't get generated until later in the process. (i.e. until `nmake` is run)

Note in particular step 4 of the "Quick start" section in the NOTES-WINDOWS file: > Use Visual Studio Developer Command Prompt with administrative privileges, choosing one of its variants depending...

If we are having new APIs for this, then we should consider how AEAD fits into it (the currently existing way of doing AEAD is fairly horrible).