Follow-up to #7338. Add fuzz coverage for the internal bundle header parser, including replacement of an owned result, cleanup after malformed input, and reuse after supported or unsupported headers. This changes no parser behavior and adds no public API.

Changes

• Add bundle_fuzzer, using the existing fuzzer discovery and standalone corpus-replay infrastructure. Each input is parsed through both the memory entry point and a callback returning at most seven bytes per read, without seeking. Both parsers initially hold a valid header, and their return codes and retained results must agree field by field.
• Check source release, cleared results on hard failures, and the two retained states for unsupported input. Complete headers must have a recognized version and object format, matching OID types, and a pack offset immediately after the blank separator with 17 <= pack_offset <= input size. Reparse a known header to check recovery, then dispose each parser twice.
• Add 35 seeds covering v2/v3, SHA-1/SHA-256, unsupported capabilities, invalid syntax and ordering, byte handling, and the 8192-byte read boundary. Mark the corpus -text -whitespace to preserve intentional byte sequences and malformed whitespace across checkouts.
• Add a deterministic test accepting a header line of exactly 1 MiB and rejecting one byte more. This checks the limit independently of the fuzzer's structural invariants.
• Correct the bundle.h comment: an unknown object format returns GIT_ENOTSUPPORTED before the remaining header is validated.
• Fix two directly relevant instructions in docs/fuzzing.md: standalone fuzzers can coexist with tests and examples, and libFuzzer builds need -fsanitize=fuzzer-no-link in CFLAGS to instrument the library objects.

Validation

Mutation campaigns ran with libFuzzer, ASan, UBSan, and LeakSanitizer using Debian Clang 19.1.7 on aarch64 Linux in Docker. A deliberately leaking test program first confirmed that leak detection worked. The campaign builds included -fsanitize=fuzzer-no-link in CFLAGS to instrument parser code.

No crashes, leaks, or timeouts were reported. The final run produced no artifacts. Mutation did not grow inputs close to the 1 MiB line limit, so boundary inputs were exercised separately.

• Under ASan and LeakSanitizer, the 35 committed seeds and the 1,130-input generated corpus replayed cleanly; all nine fuzzer replay tests and libgit2_tests -sbundle::parse passed.
• On macOS, 362 generated inputs passed under UBSan: empty input, every truncation of representative v2 and v3 SHA-256 headers, signature variants, newlines at offsets 8186–8199, and prerequisite/ref lines of 1 MiB minus one byte, exactly 1 MiB, 1 MiB plus one byte, and 2 MiB, with and without a final newline.
• Coverage inspection identified five paths reached only by mutation; corresponding seeds were added. Uncovered parser paths include read, seek, and allocation failures, the descriptor entry point, and dispose(NULL). The harness does not inject source or allocation failures.
• Standalone fuzzers and tests build with ENABLE_WERROR=ON using Apple Clang 17. A final local recheck passed all nine fuzzer replay tests and all 29 bundle::parse tests. git diff --check passes.

ASan hangs at startup on the macOS host, so ASan validation ran on Linux. Windows/MSVC validation is left to CI.