connect: the other three cluster entry points parse with no size ceiling #1107

Follow-up from #1090 by @Adityakk9031 — thanks again for that PR!

What

#1090 puts a 256 MiB ceiling on ClusterActivation before it is parsed, and rac_connect_cluster_validate_activation_proto is now guarded (core/src/connect/rac_connect.cpp:957). The other three cluster entry points that came in with #723 are not:

• rac_connect_cluster_start_proto parses at core/src/connect/rac_connect.cpp:746
• rac_connect_cluster_join_proto parses at :838
• rac_connect_cluster_stop_proto parses at :904

None is preceded by a request_size check. The only pre-parse validation any of them gets is rac_proto_bytes_validate (core/src/foundation/rac_proto_buffer.cpp:30-38), which rejects a null pointer with a non-zero size and anything above INT_MAX and nothing else. A crafted ClusterStartRequest carrying millions of ClusterLayerAssignment entries is therefore allocated in full before any cluster validation runs — the same untrusted-allocation shape #1078 named for tensor_data, at lower yield because no single bytes field is involved.

Why it matters

This is the same boundary #1090 just fixed, on the three sibling entry points in the same file, and it wants closing before the transport lands rather than after: #1071 (feat(connect): consume ClusterActivation.dtype in transport and SDK bindings (#1053)) is writing the receive path over these exact entry points, and a size check added once a wire shape is in use is one a peer can skip. Issue #1078 itself suggested the shape: "enforce it in parse_message — or a cluster-specific parse helper — before ParseFromArray".

Suggested approach

Put the ceiling where every cluster path must pass through it rather than on one function: a parse_cluster_message helper that applies the size check before ParseFromArray and routes all four entry points through it, so the limit is one constant applied at one boundary. #1090's test 11c (core/tests/test_connect_proto_abi.cpp:443-459) shows the pattern that pins the ordering without allocating the payload — a 1-byte buffer with an over-ceiling claimed size — and the same trick extends to ClusterStartRequest/JoinRequest/StopRequest.

Done when

• All four cluster entry points reject an over-ceiling request before parsing it, through one shared check.
• Each of the three new cases is covered by a test that fails if the check moves after the parse.

Not blocking #1090. @Adityakk9031, you know this code well now — you're welcome to take this one if you're interested.

Opened with help from Claude Code and Codex.

Reviewed with help from Claude Code and Codex.