blindqa 0.3.0 — no AI model in the test loop

Changed

• No AI model in the test loop. Every judgment is now a script (src/judge.mjs), so the same screen
  always gets the same verdict and a run costs no model tokens:
  • crawls check each screen for error, blank or stuck-loading pages, raw ids, codes and timestamps shown to
    users, and control names a screen reader can't tell apart (symbol-only, or the same name on every row);
  • act mode reads a control's likely effect from its role, name and attributes, a field's kind from its
    type, autocomplete hint and label, and the outcome of every click or submit from what changed on screen
    and what the server answered. Outcomes the signals can't settle are kept with "sure": false for the
    agent (blindqa_unsure) instead of being reported.
• New act-mode findings: submit-failed-silently (form closed as if saved, server refused),
  silent-refusal, crash-on-action; sign-outs after a submit are reported too.
• Journeys get the same checks as blindqa.judge.

Added

• Landing page (site/) built from 21st.dev components, pre-rendered to static HTML with structured data
  (SoftwareApplication, SoftwareSourceCode, FAQPage); docs pages restyled to match, with breadcrumbs, a 404
  page and llms.txt. "Star on GitHub" button with the live star count. CI type-checks and builds the site.
• Act mode works on one-screen apps (no navigation links): it acts where sign-in lands.

Fixed

• Act mode no longer reports "silently nothing" for a form whose submit button a person can't reach (that
  is already reported as unreachable); the press is recorded as not done.
• The screen digest left out the wrong marker, so the robot's own cursor notice could count as a change.

Removed

• The optional external judgment model and everything that used it: doctor --jev, jev-ping,
  crawl --no-jev, journey --shadow and shadow-report, MCP blindqa_shadow_report and
  blindqa_escalations (use blindqa_unsure), and the plugin's API-key settings. No keys are needed.

Upgrade: claude plugin marketplace update blindqa && claude plugin update blindqa@blindqa (Claude Code) or codex plugin marketplace upgrade (Codex). No API keys are needed any more.