Delete superseded docs and dead code

[rgarcia]

Summary

Remove artifacts left behind by the CLI, agent-class, package-collapse, and provider removals:

• delete the superseded 642-line agent tool configuration spec; current behavior remains documented in the READMEs and architecture doc
• delete unused aggregate action helpers, mode-era defaults, a zoom alias, and an uncalled Responses threading wrapper
• remove the no-op API-key test, retired Meta CI secrets, stale Meta provider prose, and the architecture reference to the removed Loop mutation API
• consolidate the duplicate smoke/matrix examples into one npm-wired Agent example and one AgentHarness example, parameterized with --model and --scenario; retain the shared provider-aware tool policy and its tests
• add the Anthropic and Google SDKs as dev dependencies so every documented native-action probe branch works after a clean install

Generic gitignore entries remain unchanged. The provider-matrix functionality is preserved through the parameterized examples rather than separate scripts.

Verification

• npm run typecheck
• npm run build --workspace @onkernel/loop
• npm test --workspace @onkernel/loop — 34 files, 429 tests
• invalid scenario smoke: exits before browser provisioning with the available scenario names

---

Note

Low Risk
Mostly documentation deletion, dead-code removal, and example/CI housekeeping with no runtime behavior changes to catalog compilation or provider streaming.

Overview
This PR finishes cleanup after earlier CLI, agent-class, and provider removals by deleting the superseded agent tool configuration spec and trimming code that nothing calls anymore (mode-era computer defaults, aggregate action schema helpers, and an unused Responses threading wrapper).

CI and docs no longer reference META_API_KEY or the retired Meta provider; architecture.md describes catalog recompilation instead of the removed setTools() API.

Examples replace separate OpenAI smoke scripts with one agent.ts and one harness.ts, driven by shared --model / --scenario parsing and the existing provider-aware toolsForModel policy (tests cover flag parsing and catalog compilation).

Dev dependencies add @anthropic-ai/sdk and @google/genai so documented native-action probe paths work after npm ci.

^{Reviewed by Cursor Bugbot for commit b8d45db. Bugbot is set up for automated code reviews on this repo. Configure here.}