Improve AI scoring and BOSS greeting reliability

[elowenzhouyb-source]

What changed

• harden DeepSeek scoring and greeting generation with larger thinking-output budgets, retries, and recoverable failure states
• keep failed scoring and greeting jobs visible and retryable in the dashboard
• make BOSS interactions run in the foreground with browser-level mouse and keyboard events
• wait for dynamically rendered communication buttons and tolerate transient page reload errors
• distinguish the first-contact composer from the preset-greeting dialog
• require a stable BOSS delivery state before recording a greeting as sent
• prevent duplicate sends when an identical delivered message already exists

Root cause

DeepSeek thinking responses could exceed the original output budget. BOSS also exposes multiple asynchronous job-detail and chat variants: background targets, delayed button rendering, first-contact dialogs, and optimistic local message bubbles could stop the workflow early or incorrectly report success.

User impact

AI failures remain recoverable, confirmed delivery follows a foreground interaction path closer to manual use, and only messages reported as delivered by BOSS are counted as sent.

Validation

• python -m unittest discover -s tests -q - 168 tests passed
• node --check src/bosshunter/browser/runtime/cdp-proxy.mjs
• git diff --check
• published diff scanned for credentials, tokens, BOSS security parameters, runtime chat content, and local company data