feat(grok): add stream command capturing full SSE response

[Daily-AC]

Summary

Adds grok stream — captures grok.com's POST /rest/app-chat/conversations/new SSE response in full and emits one row with response, thinking, model, conversationId, responseId, title, and newline-joined images. Companion / motivation: #1847.

grok ask stays untouched. The new command sits next to it for callers who want the assistant's reasoning trace, server IDs, model id, and generated-image URLs without an extra round-trip through the DOM.

How it works

1. Installs INSTALL_STREAM_INTERCEPTOR_JS once per tab (idempotent via window.__opencliGrokStreamPatched).
2. The interceptor wraps window.fetch; for any request hitting /rest/app-chat/conversations/new it response.clone().body.getReader()-drains the SSE chunks into a buffer, then pushes a {url, method, status, body, capturedAt} envelope onto window.__opencliGrokStream.
3. The Node side reuses the existing ensureOnGrok / isLoggedIn / startNewChat / sendMessage helpers from clis/grok/utils.js to drive the composer, then polls the queue until the envelope shows up (or --timeout seconds elapse).
4. summarizeResponse walks the newline-delimited JSON frames and reduces them: thinking-tagged tokens → thinking, untagged tokens → response, terminal modelResponse wins over the streamed concatenation, modelResponse.generatedImageUrls → images.

Not using Strategy.INTERCEPT / page.installInterceptor() here because the upstream interceptor calls await response.clone().json(), which silently drops Grok's newline-delimited JSON. The fetch hook for stream keeps the raw body string and parses per site — same approach we'll need for chatgpt/claude/gemini/deepseek SSE shapes in the follow-up PRs.

Test plan

• npx vitest run clis/grok/stream.test.js — 8 unit tests on iterFrames + summarizeResponse against a hand-crafted SSE body (final text, thinking trace, IDs, title, image URL list, malformed-line skipping, empty-body safety)
• npx tsc --noEmit clean
• npm run test:adapter — 380 files / 3722 tests pass with the new file added
• npm run build — manifest regenerates, dist/src/main.js exposes opencli grok stream
• Live E2E against grok.com:

  $ node dist/src/main.js grok stream "say only pong" -f json
  [
    {
      "response": "pong",
      "thinking": "Thinking about your request",
      "model": "grok-3",
      "conversationId": "00c9df44-9f05-4002-8914-fb708d9bacd2",
      "responseId": "dde78767-3bef-4e06-864e-7ae44a566cda",
      "title": "pong",
      "images": ""
    }
  ]
  

Follow-ups

If this lands cleanly I'll open separate PRs for chatgpt stream, claude stream, gemini stream, deepseek stream — each parser shape differs enough to deserve its own diff. Detail of each shape is in the issue.