Accept newly downloaded input files on replay continuations #370

A programmatic script can pause for an MCP tool that downloads a file after execution has already started. The continuation currently accepts tool text results but cannot add the newly registered input to the execution's file set, so reading the file in the same script fails.

Related host reproduction and external testing: LibreChat-AI/LibreChat#17092

I would like to contribute an additive replay-continuation field for owned input references, associated with issued tool-call IDs. The API would validate the references and merge them under the execution lock before replaying user code, with immutable/idempotent retries and existing limits. Pending replay responses would advertise the capability; blocking-mode continuations would remain explicitly unsupported.

Synthetic Python and Bash same-script reads have been exercised through a real local Code API/NsJail stack. The host and agents SDK changes are separate contributions. Happy to take this issue; please assign it to me if assignment is preferred here.