Productionize Agent Browser Driver: auth-state reuse, modal reliability, CDP perf path, and critical-flow tests

[drewstone]

Summary

This PR consolidates all in-flight work into a single production-ready batch:

• Auth/session reliability

  • Add CLI + config support for pre-authenticated sessions via --storage-state / storageState
  • Add scripts/save-storage-state.mjs and pnpm auth:save-state for one-time manual login capture
  • Fix modal interaction regression: actions now attempt first, then dismiss blocking overlays only on failure and retry once

• Observation performance

  • Add CDP fast-path for accessibility snapshot + page metadata
  • Export CDP helpers for reuse
  • Keep Playwright fallback path intact

• Runner/artifact robustness

  • Improve artifact sink behavior for Playwright video flush edge cases
  • Harden test-runner reliability around artifact/progress handling and critical failure paths

• Benchmarks + fixtures

  • Add observe and full-agent benchmark scripts plus deterministic fixture pages

• Coverage

  • Add/expand integration and critical-flow tests for:
    • modal click behavior
    • CDP snapshot + metadata path
    • preview verification
    • parallel worker timeout/deadlock handling
    • artifact persistence behavior

• Repo hygiene

  • Ignore local auth + benchmark output directories (.auth/, audit-results/, bench/results/)

Why

• Enables realistic E2E usage against apps with auth without sharing raw secrets/cookies
• Fixes a concrete blocker seen on real-world auth modals
• Improves per-turn observation performance while preserving compatibility
• Raises confidence with non-mocked critical-path tests

Validation

• pnpm build
• pnpm test (13 test files, 123 tests passing)
• Real runs against https://ai.tangle.tools using agent-browser-driver
  • After fix, driver successfully executes auth modal steps (Sign in -> Other sign-in options -> Sign up)
  • Flow then correctly aborts on product-side email verification gate (expected external blocker)

Notes for Review

• This is intentionally one consolidated PR per request; scope is broad but internally cohesive around reliability/perf/testability.