Deactivate DeepSeek/Anthropic A/B round-robin for article simplification

[mircealungu]

Why

The provider A/B experiment (round-robin between DeepSeek and Anthropic for article simplification, to compare error rates) has run long enough. Deactivate it and pin providers deliberately.

Current behavior

zeeguu/core/llm_services/simplification_and_classification.py:

• _get_next_simplification_provider() (~L27) alternates 'deepseek' / 'anthropic' on a module-level counter.
• It's used by simplify_article_adaptive_levels() (~L79) only when the caller passes no simplification_provider (~L122-127).
• The batch crawler already pins DeepSeek — zeeguu/operations/crawler/crawl.py:148 crawl_round_robin(..., simplification_provider='deepseek') — so batch never round-robins.
• The remaining round-robin caller is the on-demand path: simplify_article_adaptive_levels(title, content, language_code) at ~L454 (creating a simplified level when a reader needs it), which passes no provider.

Change

Stop the on-demand path from alternating; pin it to Anthropic (Haiku), matching models.py SIMPLIFICATION = ANTHROPIC_HAIKU ("real-time Haiku key path"). Net result:

• On-demand / live simplification → Anthropic Haiku (low latency; DeepSeek stays as fallback on failure).
• Batch crawl simplification → DeepSeek (already pinned; cheaper; latency-insensitive).

Minimal options (pick one):

1. Set the default in simplify_article_adaptive_levels so simplification_provider defaults to 'anthropic' instead of round-robin, and remove/retire _get_next_simplification_provider.
2. Or make _get_next_simplification_provider() return a fixed 'anthropic' (leaves less dead code behind — prefer removing it).

Acceptance criteria

• On-demand simplification deterministically uses Anthropic Haiku (verify via logs / result['provider']).
• Batch crawl still uses DeepSeek.
• Round-robin counter and helper removed (no dead code).
• DeepSeek fallback-on-failure behavior preserved.

Notes

• The related grammar-correction experiment is already parked (correct_grammar=False by default, ran 2025-12-09 → 2026-04-05) — out of scope here, but confirm it stays disabled.
• Context: this cleanup also sharpens a production example in the "LLM integration patterns" paper (live=Anthropic / batch=DeepSeek as a clean latency-tiered split).