fix: use LLM for unique video extension prompts; remove tokens on 400 email-domain-rejected

[crowwdev]

PR Title

fix: use LLM for unique video extension prompts; remove tokens on 400 email-domain-rejected

---

Summary

Two independent fixes:

1. Video extension scene repetition — Each extension round now gets a unique LLM-generated continuation prompt instead of reusing the original prompt, preventing repeated scenes in long videos.

2. Auto-remove banned tokens — When a token refresh returns 400 with email-domain-rejected in the response body, the token is immediately removed from the pool.

---

Changes

• Bug fix
• Feature
• Refactor
• Documentation
• Test

app/services/grok/services/video.py

• Added _generate_continuation_prompt() using curl_cffi.requests.AsyncSession (no new dependencies)
• Reads app.api_key from config and sets Authorization header if configured
• Sets "stream": False explicitly in the LLM request
• Uses grok-3-fast model with temperature=0.8, max_tokens=60
• Falls back to original prompt silently if LLM call fails
• _stream_chain() and _collect_chain() call LLM for each extension round (plan.is_extension == True)

app/services/grok/services/video_extend.py

• Imports _generate_continuation_prompt from video.py

app/services/token/manager.py

• In record_fail(): when status_code == 400 and reason contains email-domain-rejected, token is immediately removed from pool
• In _refresh_one(): same check during cooling token refresh — 400 + email-domain-rejected removes token from pool instantly

---

Verification

• 30-second video (5 rounds) generates 5 distinct prompts visible in logs
• Tokens with email-domain-rejected disappear from pool immediately after next refresh cycle
• Container starts without errors (no new pip dependencies)
• Fallback to original prompt works when LLM is unavailable