[m-messer]

Summary

• Add EvaluationFunctionError base class; InvalidSubmissionError now subclasses it, plus new UpstreamTimeoutError and UpstreamUnavailableError. All exported from lf_toolkit.
• Each class sets its JSON-RPC code (rpc_code) and file IO code (file_code):
  • InvalidSubmissionError → 422 / INVALID_SUBMISSION
  • UpstreamTimeoutError → 424 / UPSTREAM_TIMEOUT
  • UpstreamUnavailableError → 424 / UPSTREAM_UNAVAILABLE
  • EvaluationFunctionError → 500 / INTERNAL_ERROR
• Over JSON-RPC the file code is also sent as the error's data ({"code": ...}), so the two 424s can be told apart. The message reaches the caller as-is, without the "Error calling user handler" prefix.
• Upstream failures use 424 (Failed Dependency), the same as shimmy, because the function itself is working.
• Stream server logs no longer print request/response contents or exception messages (these can contain student submissions). They print only sizes, the exception type and the traceback.
• README: new "Reporting Errors" section.

Test plan

• New/updated tests in tests/io/ (handler, rpc_handler, file_server, stream_io); 213 passed locally
• CI green

🤖 Generated with Claude Code