Groundwork to restrict where credential-helper settings can come from (inert in this build)

What

Claude Code now contains a filter for credential-helper settings. A credential helper is a setting that names a command Claude Code runs to fetch a key, a login or a header. The filter covers apiKeyHelper, awsAuthRefresh, awsCredentialExport, gcpAuthRefresh, otelHeadersHelper and proxyAuthHelper.

• Every settings source is now passed through the filter before the sources are merged. The filter can remove these keys from a source that is not allowed to supply them.
• New helpers decide whether managed policy settings (policySettings, the settings an administrator deploys) are treated as machine_admin or as_read. The choice depends on whether the managed source can be written by an ordinary user.
• In machine_admin mode, the helper keys are removed from the other sources and re-read only from admin-only sources. Remotely delivered settings and MDM sources (device-management profiles) that a user can write to are left out.
• A helper is disabled if it was armed from a remote source or from a user-writable base. The new helperArmedFromUserWritableBase check tracks the second case.
• In this build, the switch that picks the mode always returns as_read. The live settings loader also hard-codes credentialHelperKeysFrom: () => "as_read". As a result, no keys are removed yet.

Why

Nothing changes for you today, because settings are still merged exactly as read. The code is built so that, once it is switched on, a credential helper command can come only from managed settings an administrator controls, and not from a settings file an ordinary user can edit.