fix(browser-worker): per-conversation browser session isolation — cookies/login state bleed across conversations via shared navPage #11539

Part of umbrella #11536 (OpenManus research adoption, Task 3).

Problem

The browser worker holds a single global navPage (autobot-browser-worker/playwright-server.js, let navPage = null) — one browser, one context, one page for ALL conversations and users. Cookies and login state from one conversation are visible to the next: log in to a site in conversation A, and conversation B (potentially another user) browses with those credentials. This is session bleed, not just a missing feature.

Per-session contexts already exist in autobot-backend/research_browser_manager.py (ResearchBrowserSession(session_id, conversation_id) — dedicated context per session) but the chat workflow path does not use them; send_to_browser_vm() (autobot-backend/api/browser_mcp.py:543-583) talks to the shared page.

Proposal

• Browser worker: map session_id → Playwright BrowserContext (one browser process, N isolated contexts), with idle GC mirroring the docker workspace pattern (services/docker_task_workspace.py idle expiry).
• send_to_browser_vm() passes the conversation/session id on every call.
• Reuse/converge with research_browser_manager.py session semantics rather than inventing a parallel mechanism (canonical-source rule).

Acceptance Criteria

• Two concurrent conversations get isolated cookie jars (test: set cookie in A, assert absent in B)
• Contexts are GC'd after configurable idle timeout
• Existing single-session behavior unchanged for a lone conversation
• No second session-management implementation left behind (converge with research_browser_manager)