feat: add a usable durable Harness goal interface

[moortekweb-art]

What changed

• adds a first-class Harness page with a free-form goal prompt
• supports the configured managed worker and arbitrary OpenAI-compatible endpoints
• adds durable start, poll, continue, approve, and stop interactions
• surfaces progress, review requests, terminal outcomes, verification evidence, and changed paths
• adds server-side Harness proxy routes and focused model/proxy/navigation tests

Why

The existing Harness integration exposed routing and operational state but did not give a new user a complete place to enter a goal and work with it until completion. This makes the product usable as an interactive goal loop rather than only an operator/status surface.

User impact

A user can type a plain-language objective or optional /goal command, choose the managed worker or an OpenAI-compatible provider, and follow the task until it completes or needs a decision.

Validation

• 26 focused frontend tests passed
• ESLint and Prettier completed in the commit hook
• TypeScript typecheck passed
• gitleaks scanned all five new commits and found no leaks
• live Node1 deployment was exercised through the Harness page

Draft review notes

• review generic product wording for the legacy live-route registry, which still describes Node1 primary and Node2 overflow
• keep environment-specific endpoints, credentials, workspace paths, and Tailscale details outside the upstream patch