[bug] Server collapses JSON extractor errors to HTTP 400

[ting-hong-shieh]

Symptom

The Rust server converts every Axum JsonRejection into the same HTTP 400 invalid_body response. This changes status codes that clients use to distinguish an unsupported media type from an oversized request.

On current main:

Request	Expected	Actual
Valid JSON without Content-Type: application/json	HTTP 415	HTTP 400 invalid_body
Body larger than the configured 32 MiB limit	HTTP 413	HTTP 400 invalid_body

Reproduction

Start the server with any valid route configuration:

switchyard-server --config routes.toml --host 127.0.0.1 --port 4000

Missing content type:

curl -sS -o response.json -w '%{http_code}\n' \
  http://127.0.0.1:4000/v1/chat/completions \
  --data '{"model":"noop-route","messages":[]}'

Observed:

400
{"error":{"code":"invalid_body","message":"Request body must be valid JSON: Expected request with `Content-Type: application/json`","type":"invalid_request_error"}}

Oversized body:

mkfile 33m oversized.bin
curl -sS -o response.json -w '%{http_code}\n' \
  http://127.0.0.1:4000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  --data-binary @oversized.bin

Observed:

400
{"error":{"code":"invalid_body","message":"Request body must be valid JSON: Failed to buffer the request body: length limit exceeded","type":"invalid_request_error"}}

Expected vs. actual

• Expected: Preserve the extractor rejection status: 415 for a missing JSON content type and 413 for a body over the configured limit. Anthropic requests should render the existing 413 request_too_large error type.
• Actual: llm_json_body stringifies the rejection, then invalid_body_error always creates a 400 response. The server's explicit Anthropic mapping for StatusCode::PAYLOAD_TOO_LARGE is therefore unreachable for this extractor failure.

Environment

• Switchyard commit SHA: b256d936f1d77bf13ec9bec399ea0a253e07ca05
• Rust version: rustc 1.96.1 (31fca3adb 2026-06-26)
• OS / arch: macOS 26.5.2, arm64
• Install path: source checkout
• Inbound format: OpenAI Chat Completions; the same extractor path is shared by all three endpoints
• Backend: local noop route; no provider credentials

Additional context

The relevant path is crates/switchyard-server/src/lib.rs: endpoint extractors return Result<Json<Value>, JsonRejection>, llm_json_body reduces an error to a string, and invalid_body_error hard-codes StatusCode::BAD_REQUEST.

I searched open and closed issues and pull requests for JSON extractor errors, length limit exceeded, missing content type 415, payload too large 413, and invalid_body status; I found no matching report or in-flight fix. PR #320 is related error-envelope work but does not preserve extractor status codes.

I'm happy to send a PR with focused 400/413/415 regression tests if this behavior should be preserved.