v1.12.1

A controller patch release over 1.12.0. Most of its fixes concern the page a navigation the gate stopped leaves
behind, Chrome's error page (chrome-error://chromewebdata/). One closes a DevTools session leak that any page could
hit. The client, LangChain and MCP bridge packages change only in version; update the controller image to get it.

Upgrade notes

• Status codes that changed. Over REST, a fork without start_url and a shadow browse from Chrome's error page
  answer 400 (1.12.0: 403 and 500). Over MCP, both answer a fixed error with no code (1.12.0:
  not_permitted). A shadow browse from a refused, data: or blob: page answers 403 (1.12.0: 500).
• A new key, absent unless a navigation was stopped. blocked_navigation can now appear on observe,
  browser.get_html and an action result's after observation, beside the tab list's.

Fixed

• Fork and shadow browse from Chrome's error page. A session whose last navigation the gate stopped, or that
  failed, sits on chrome-error://chromewebdata/. POST /sessions/{id}/fork without start_url, and
  POST /sessions/{id}/shadow-browse, now answer 400 before exporting any state, and detail says the page is
  Chrome's error page for a navigation that was stopped or failed; for fork, it says to pass start_url, which
  works as before. In 1.12.0, fork exported the session's state and then answered 403 "Not permitted", and shadow
  browse answered an unhandled 500: a REST client that branches on the status now sees 400 where it saw 403 or
  500. Over MCP, browser.fork_session and browser.enable_shadow_browse answered
  {"error": "URL scheme 'chrome-error' is not allowed; only http and https are", "code": "not_permitted"}; they
  now answer the fixed text as error, with no code.
• Shadow browse over REST from a refused, data: or blob: page answered an unhandled 500 with a traceback.
  It answers 403, as fork does: a page ALLOWED_HOSTS refuses gets the page routes' detail, any other refusal
  "Not permitted".
• The built-in agent loop saw no label. /agent/step, /agent/run and the queued and cron runs prompt the
  model with a compact observation that dropped blocked_navigation, so on Chrome's error page the model saw the
  page with no word on what was stopped. The compact observation keeps it.
• A DevTools session left attached after a slow target lookup. The read of a page's target id is bounded. When
  the bound fired while the browser was still opening the session the read uses, the session it then opened stayed
  attached for the life of the page. It is now detached once it opens.

Changed

• The error page says what was stopped. While a session's page is Chrome's error page for a navigation the gate
  stopped, observe (GET and POST /sessions/{id}/observe, MCP browser.observe), browser.get_html and an
  action result's after observation (so the actions.jsonl record and the witness receipt built from it) carry
  blocked_navigation: the origin of the stopped request, the reason and an explanation, as the tab list did in
  1.12.0. The error page's own DOM is returned as before; it carries the refused URL in full, path and query
  included, since the page chose that URL. The label goes once a later request of the page continues, whatever
  became of it. The tab list decides it on the URL it lists, as 1.12.0 did, so a tab listed as allowed never
  carries one; observe takes the active tab's, so the two agree and the block is looked up once per observation.