Organization policy loading records more failure detail, and file attachments wait for it

What

In organizations that manage Claude Code centrally, Claude Code fetches the organization's policy limits and remote settings from a server. This release changes what happens when that fetch fails or has not finished yet.

• Attaching files now fails with the error code policy_pending while the organization policy has not loaded. The message tells the model to try again in a minute and, if it is refused again, to tell the user. A digest_mismatch size and digest check used to sit at this point.
• The policy-state reason now carries policy_server_error_code and policy_server_auth_denial, and the result includes httpStatus.
• Loading policy limits can now run as a single attempt (singleAttempt, which runs as accelerated_retry). It skips the retry that has no verdict when the trigger is api_refusal_trigger.
• The startup load now records the server error code, whether access was denied, and the last attempt (startupLoadServerErrorCode, AuthDenial, LastAttempt). Before, only the server error type was recorded.
• Remote settings fetches also record the server error code and auth denial when they fail or get no response.
• Refreshing the cached policy now passes along how old the cache is.

The related setting tengu_rustling_orbit was read as off for this site's account and for the anonymous baseline, and no reading has been taken under this release yet.

Why

If you attach a file before your organization's policy has arrived, you get a clear retry message instead of a generic failure. The extra error detail mostly helps admins and support work out why a managed policy fetch failed.