fix: harden HTML adapters with live E2E coverage

[lennney]

Summary

• Harden Bing, Baidu, and Yandex HTML adapters against parser drift: recognized cards that stop parsing now report parse_error instead of silently returning an empty success.
• Classify anti-bot challenge bodies and redirect URLs before ordinary HTTP status mapping, including HTTP 200, 403, and 429 responses.
• Preserve adapter failures through both parallel and waterfall free_search execution as structured partialFailures.
• Make live E2E probes bounded and deterministic: budget exhaustion skips cleanly, a bot challenge stops later probes, provider coverage is tracked separately from extraction, and invalid engine selections fail before build/network work.

Validation

• npm run build ✅
• npm run lint ✅
• npm test ✅ 77 test files, 788 passed, 5 skipped
• git diff --check ✅
• Offline stdio E2E ✅ initialize/tools-list
• Live runner validation ✅ invalid/empty LIVE_E2E_ENGINES exits with code 2 before build/network work

Live E2E

The opt-in live suite remains bounded by LIVE_E2E_MAX_REQUESTS and LIVE_E2E_MIN_INTERVAL_MS. The existing live run was intentionally opt-in and observed provider-dependent availability: Bing returned results, Baidu returned bot_challenge, and later probes were skipped. The latest local verification did not make additional real network requests.

Live provider availability varies by exit and time; the suite does not claim stable search availability or quality.