feat(testing): run the golden cookbook suite without an API key

[s-alexey]

The golden tests were the only end-to-end coverage of the harness — tasks, schema parsing, tool loops, chatrooms, assertions — and every one of them needed a live model. They could not run on a branch, so that coverage never gated a change, and each task definition was locked inside the single 1500-line test module.

Add a public ScriptedLLM that replays canned responses at the invoke() boundary. Because it plugs in below respond(), a scripted run still goes through the real path: schema parsing, assertions, chat history and serialization all behave as they would against a live backend, just offline and deterministically. tests/mocks.py's MockedChat becomes a thin alias.

Split the cookbook into one file per theme. Each file holds its tasks, and under each task the two tests that exercise it: a scripted one that runs anywhere, and a live one parametrized over a model pool. Tests asserting a failure are scripted only — a real model may legitimately answer correctly. Scenarios whose input needs the network, or that verify provider wiring, stay live-only: faking those would assert nothing.

Model pools come from a table where each model is named once alongside what it can be used for, replacing per-feature name sets that had drifted out of sync. A pool is empty when no provider is configured, so live tests skip on an empty parameter set and the suite stays green without credentials — no marker or skip machinery.

Wire the suite into CI, and run it in TESTING mode: the default client serializes every task and run to JSON in the working directory, which left 183 files behind and dominated the runtime.

Offline: 43 tests, no credentials. Live selection is unchanged from the previous suite at 476 model-parametrized tests.