brw v0.2.0

brw v0.2.0

A semantic-browser-control release focused on making the tool faster, more
accurate, and easier for small models to use — plus forward-looking WebMCP
and an opt-in navigation guardrail.

The Chrome extension/plugin is unchanged and stays at 0.1.0. Only the
daemon bumps to 0.2.0.

Small-model ergonomics

• Actionable error recovery — ref not found / not actionable now tell the
  model to re-snapshot or inspect, instead of dead-ending.
• brw_observe in the core tool profile — the cheap "what changed" check is
  now in the lean surface.
• brwd --print-system-prompt + docs/agent-guide.md — an opinionated,
  token-efficient operating loop to prepend to an agent's system prompt.
• brw_snapshot format:"compact" — one terse line per element
  (e17 button "Submit") instead of JSON; markedly fewer tokens, same refs.

Accuracy / coverage

• Closed shadow DOM piercing — closed-root web components (many design
  systems) now show up as normal refs (document-start + in-walker install).
• Cross-origin iframe signal — unreachable cross-origin frames are surfaced in
  metadata.cross_origin_frames (box + origin) so the agent falls back to
  brw_screenshot + brw_click_xy instead of going blind.

Forward-looking / safety

• WebMCP (--enable-webmcp) — brw acts as the agent-side
  navigator.modelContext runtime; brw_page_tools lists and
  brw_call_page_tool invokes a cooperating site's declared tools.
• Navigation guardrail (--blocked-domains / --allowed-domains) — gates
  brw_open / brw_open_incognito / brw_replay_request so a prompt-injected
  agent cannot steer the browser off-limits. Opt-in, off by default.

Full test suite green (incl. headless-Chrome integration tests for shadow,
cross-origin, and WebMCP).