Add binary E2E coverage and fix exposed runtime regressions

[CuriosityOS]

What and why

Adds a standalone black-box suite that accepts a BetterWright executable or tests betterwright on PATH. The suite imports no target source/package internals, so it can exercise a built checkout or a separately installed release.

• 100 independently isolated cases cover CLI input/configuration, browser actions and observations, sessions/profiles, artifacts/recording, network and sandbox restrictions, credentials, MCP, local mock-model agent execution, stealth, provider fallback, timeouts, and live view.
• Each case uses a fresh home/profile/vault and local fixtures. Inherited credentials and preload hooks are excluded; synthetic secrets are redacted from diagnostics. No real cloud accounts, native personal cookie stores, installers, or paid services are used.
• Adds command/case deadlines, cleanup, group/filter selection, --require-all, and atomic JSON reports with explicit failures, skips, incomplete runs, and coverage limits. The existing Node tests job runs the strict suite and uploads its report.
• Documents usage and limitations in docs/binary-e2e.md; the suite complements rather than replaces unit, type, platform, and Electron tests.

bun run test:e2e
bun run test:e2e /absolute/path/to/betterwright
bun run test:e2e --binary bun --binary-arg "$PWD/dist/bin/betterwright.js" --require-all

Regressions exposed and fixed

• human.type(..., { clear: false }) now positions the caret at the end for normal typing as well as fallback insertion.
• History navigation waits for completed document readiness after commit, including BFCache restores that do not emit a new DOMContentLoaded/load event.
• MCP explicitly observes stdin EOF/close and closes both its transport and browser instead of depending on an SDK close callback that EOF never invoked.
• Live-view warm-up preserves setContent pages whose URL remains about:blank.
• The stealth worker directly selects patchright-core; Bun 1.4's handling of Node loader registration no longer silently leaves the stock driver active.

Verification

• Standalone copy of tests/e2e/, run outside the checkout against betterwright on PATH with Bun 1.4.0 and --require-all: 102 passed, 0 failed, 0 skipped (100 cases plus 2 prerequisite checks).
• Related existing real-browser regressions: 17 passed, 0 failed.
• Focused harness, navigation, MCP, human-input, and stealth unit tests: 72 passed, 0 failed.
• Credential redaction case additionally passed three consecutive synchronized runs.
• bun run lint, bun run typecheck, bun run test:types, bun run build:harness, bun run check:versions, and bun run check:build pass. Lint retains existing warnings.
• Final rebuild produced byte-identical dist/ files to those exercised by the standalone run.

Public API declarations and dependency pins are unchanged. Required CI display names are preserved; the new artifact action is SHA-pinned. No test imports the worker entrypoint. The separately requested local worker-refactor plan is excluded from Git and is not part of this PR.

[greptile-apps]

Confidence Score: 5/5

The reviewed runtime changes have focused execution evidence for browser history and editable-control behavior, and no confirmed merge-blocking issue remains.

What we checked:

• Exercised the MCP server startup with input already closed and verified that the server cleanup and browser cleanup handlers ran, and the process exited successfully.
• Built the browser validation distribution and exercised same-document navigation with BFCache restoration, confirming readiness states completed and no timeout occurred.
• Rendered editable controls with carets at the beginning and ran the typing harness with clearing disabled, confirming text was appended to all target controls.
• Validated the post-connection setup and observed server-close and browser-close events with exit code 0, noting that a real StdioServerTransport is required for a complete stdin EOF event.
• Ran the focused same-document BFCache validation and observed ok:true, readyState 'complete', and nonzero navigation timings with timeout 0.

Summary

• Adds a standalone black-box binary E2E framework and runs its strict suite from CI with uploaded reports.
• Updates editable-control typing, browser-history readiness, MCP shutdown cleanup, populated about:blank live-view handling, and stealth driver selection.
• Focused browser checks confirmed history traversal and end-append typing behavior in real Chromium. MCP cleanup was confirmed when standard input was already closed before connection.

No issues requiring changes before merging were found.

_{Reviews (1) · Last reviewed commit: "Keep E2E additions under Unreleased afte..."}