Rewrite Prime CLI around Pydantic configs

[xeophon]

Overview

This PR rewrites Prime CLI around a small command registry and strict Pydantic leaf configs, while drawing a clear boundary between Prime's platform control plane and Verifiers' local runtime.

High-level changes

• Replaces Typer routing with a data-driven registry and longest-path dispatcher.
• Gives every Prime-owned leaf a strict Pydantic Config model and run(config) entrypoint.
• Uses explicit paths such as prime eval run, prime eval submit, prime train run, and prime lab view.
• Keeps Rich responsible for output and prompts without coupling interaction to argument parsing.
• Makes Prime responsible for account contexts, Hub CRUD and publishing, public/private acquisition, downloads, wheel caching and installation, hosted evaluation and training, Lab, compute, and account workflows.
• Delegates local eval, environment initialization and validation, serving, and GEPA to Verifiers with unchanged native argv.

Command architecture

A central registry maps complete command paths to leaf modules. The root dispatcher owns only global context, version, plain mode, and root/group help. Prime-owned leaves parse with prime-pydantic-config; Verifiers-owned commands execute through the selected workspace interpreter.

Deprecated aliases, default-command groups, Typer contexts, and implicit routing are removed.

Prime and Verifiers boundary

Hub packages are acquired explicitly:

prime env install owner/environment@version
prime eval run environment

Prime resolves the selected account and passes explicit environment variables to the Verifiers process. Verifiers receives an installed local module and remains responsible for its run configuration, plugin loading, execution, lifecycle semantics, and native artifacts.

Companion change

Verifiers CLI and local-runtime surface: PrimeIntellect-ai/verifiers#1857