[afonsojanu]

yaml_parser_initialize() and yaml_emitter_initialize() both memset their object to zero before doing any fallible work, which is what lets a caller safely pass the object to the matching _delete() function even after a failed _initialize() call. yaml_document_initialize() was missing this: when it fails partway through (for example when a supplied tag directive isn't valid UTF-8), it returns 0 without ever writing to *document.

A caller that follows the same pattern used elsewhere in this library and around the codebase, initialize then unconditionally delete regardless of the return value, ends up calling yaml_document_delete() on a document that still holds whatever was previously on the stack. That's exactly what the OSS-Fuzz dumper harness for this project does, and it's how #312 reproduces: yaml_document_delete() dereferences the leftover nodes.start/nodes.top from an earlier stack frame and crashes trying to free it.

I reproduced this with ASan/UBSan using the exact commit and POC attached to #312, and it still crashes the same way on current master:

AddressSanitizer: SEGV on unknown address 0x...
    #0 ... in __asan::Allocator::Deallocate
    #1 ... in free
    #2 ... in yaml_free
    #3 ... in yaml_document_delete api.c:1131
    #4 ... in LLVMFuzzerTestOneInput libyaml_dumper_fuzzer.c:308

With tracing added to the harness, the crash happens right after copy_document() fails on the very first document (yaml_document_initialize() rejects a tag directive with invalid UTF-8), and the fuzzer's cleanup loop then deletes that never-initialized document. Adding the same memset() the other two _initialize() functions already use fixes it: after the change the exact same input runs clean under ASan/UBSan.

Added tests/test-document-initialize.c, which builds a tag directive with an invalid UTF-8 prefix to force the failure path, then asserts the document comes back zeroed and that deleting it doesn't crash. It's wired into both tests/Makefile.am and tests/CMakeLists.txt alongside the existing tests. I ran it both ways: it fails (assertion on the memcmp) without the fix and passes with it, and the full CMake test suite (version, reader, nesting, document-initialize) passes.

Fixes #312.

[afonsojanu]

Closing this one too, same reason: starting fresh and clearing out everything I had open.