The /mcp nest (to_mcp streamable HTTP service) collects the raw body
itself, so axum's extractor-based DefaultBodyLimit never applies and a
multi-GB chunked POST buffers entirely in memory (review-001 SRV-03).
Wrap the nest with an explicit counting-body middleware: 8 MiB cap
(headroom over the gateway's 2 MiB for JSON-RPC batch payloads),
Content-Length-declared oversizes rejected before reading, streaming
oversizes cut off mid-read and answered 413 (rmcp maps body-read errors
to 500, so the middleware sources the status itself).
Verification: cargo check (default + mcp + all-features); mcp test
suite pending tree unstuck from parallel agent's WIP.
glm-5.3-flash
a943d142c4