feat: wire Portal drawer to in-browser agent runtime

[w1kke]

Summary

• integrate OpenClaw Lite-style browser agent path into Portal’s existing Agent drawer
• move drawer chat execution to in-browser runtime flow (no separate OpenClaw process)
• add same-origin LLM proxy endpoints:
  • POST /api/llm/openai/v1/chat/completions
  • POST /api/llm/openai/v1/responses
• add runtime capability endpoint:
  • GET /api/runtime/capabilities
• support optional Codex CLI local bridge (OPENCLAW_LITE_CODEX_CLI=1) via server/modules/agent/codex_bridge.js
• keep /api/agent/chat as explicit guardrail (BROWSER_AGENT_ONLY)

Hatching integration

• add explicit hatching start endpoint:
  • POST /api/hatching/start
• extend session state with hatching lifecycle fields (started, startedAt, completedAt, mode)
• add onboarding control in Portal UI:
  • Start hatching with browser agent button
  • hatching status indicator
• when hatching starts, the chat window opens and the browser agent + user progress onboarding together with phase hints:
  • connect agent
  • match sigil
  • press open
  • complete

UI/test additions

• drawer key UX for OpenAI key (stored in browser localStorage)
• test-only GET /__test__/llm/stats
• e2e coverage:
  • e2e/08_agent_browser_chat.spec.js

Why

Per product direction: the Portal chat agent is the browser agent itself (no other OpenClaw runtime dependency), and hatching should start and run collaboratively inside Portal.

Verification

• npm test -- e2e/08_agent_browser_chat.spec.js
• npm test -- e2e/01_home.spec.js e2e/02_match_unlock.spec.js e2e/03_create_share_leaderboard.spec.js e2e/04_referral.spec.js e2e/05_erc8004_mint.spec.js e2e/06_agent_house_append.spec.js e2e/07_token_house.spec.js