browser: support/verify interactive `browse` driver mode inside the VM

[NathanFlurry]

The registry browse CLI's interactive driver mode (browse open, browse click, browse fill, …) is currently documented as requiring a sandbox. Per the bundled skill (examples/browserbase/skills/browse/SKILL.md), driver commands auto-start a background daemon that holds the live session (active page, tabs, element refs) between commands and talks to each invocation over a local socket.

This should plausibly work inside the VM — the kernel has a full process table and background processes — but it is currently unverified:

• The docs claim it is unsupported, inherited from the original page copy.
• An attempt to verify e2e locally (browse open then browse status across separate agent.exec() calls) was blocked before reaching browse: the local dylib-actor boot path needs the r6 sibling engine build (packages/agentos/tests/actor.test.ts gates this behind AGENTOS_E2E_FULL=1 and calls it env-fragile).

Acceptance

• E2E test: browse open <url> in one exec, browse status in a second exec sees the same session (daemon persisted), interactive commands drive a cloud session from inside the VM.
• If the daemon needs runtime support (detached long-lived guest processes + local socket reuse across execs), implement or scope it.
• Update website/src/content/docs/docs/browser.mdx (Interactive browsing note links here) once supported.

🤖 Generated with Claude Code