Revisit structured errorCode for sandbox lifecycle API errors

There was a previous closed PR for the base mechanism:
https://github.com/e2b-dev/infra/pull/2457

It added optional structured errorCode support to API errors, but it was closed without merge and current main does not appear to include it. I would like to check whether this direction is still acceptable, especially for sandbox lifecycle errors.

Today, different sandbox lifecycle failures can collapse into the same HTTP status and human-readable message. For example:

{
  "code": 500,
  "message": "Error killing sandbox: sandbox operation failed"
}

For clients, message is hard to use safely. It is intended for humans and can change with wording, wrapping, or lower-level error formatting. But clients need a stable way to decide whether an error is retryable.

Some sandbox errors should not be retried, while others may be transient:

• sandbox_abnormally_exited: sandbox exited unexpectedly; retrying kill is not useful because the sandbox process is already gone
• sandbox_state_conflict: delayed retry may be appropriate
• sandbox_node_unavailable: bounded retry may be appropriate
• sandbox_delete_rpc_failed: final state is unknown, retry briefly then escalate

Could E2B API errors expose an optional machine-readable errorCode, following the direction of #2457? It would not need to be present on every error response; handlers could omit it until a stable semantic code exists.

Example:

{
  "code": 500,
  "errorCode": "sandbox_abnormally_exited",
  "message": "Error killing sandbox: sandbox operation failed"
}

This would keep code and message backward compatible, while allowing SDKs and clients to avoid parsing human-readable messages.

Start by reviewing closed PR #2457 and the current sandbox lifecycle error handlers to compare the proposed base mechanism with main. Trace how lifecycle failures become HTTP responses, then determine which errors have stable retry semantics and what compatibility checks or tests would be needed for an optional errorCode design.