claude plugin eval reaches general availability with a trust-confirmation requirement

What

• claude plugin eval now reports itself as generally available for every client type (Bedrock/Vertex/Foundry, gateways, telemetry-disabled clients, CI), unless a server-side flag turns it off — in which case the CLI now says the command exists but is switched off, rather than claiming it doesn't exist.
• Before loading a target, claude plugin eval now runs a trust check that must resolve to trusted: true — via an already-installed plugin, an existing trust marker, the --trust-plugin flag, or an interactive prompt — or the command errors out and exits without loading or running anything.
• The command's help text was expanded to explain that it loads and runs a plugin's eval suite on your machine, that sandboxing limits but doesn't guarantee safety, and that the first run against an untrusted plugin directory will prompt for confirmation (use --trust-plugin to pre-answer that, e.g. for CI).
• claude plugin eval init, when run without a terminal but detected as launched from inside another Claude Code session, can now delegate its authoring interview to that parent session instead of failing with "no TTY available". In that delegated interview flow, the interviewer must now explicitly ask whether you trust the plugin directory before piloting eval cases, and only adds --trust-plugin on an explicit yes; case files are still written on a no/no-answer, just not piloted.

Why These changes make claude plugin eval a supported, documented feature everywhere while making sure a plugin's eval code — which runs on your machine — isn't executed against an untrusted directory without your explicit confirmation.