Auto-fetch ELO scores instead of hardcoding in generate_model_catalog.py

[gltanaka]

Problem

generate_model_catalog.py hardcodes ELO scores in a static ELO_SCORES dict. This causes several issues:

• Scores go stale — ELO scores drift as new votes come in on the Arena leaderboard, but our dict stays frozen
• New models are invisible — models not in the dict get no score and may be filtered out
• Multiple leaderboards — there are different coding leaderboards (Code Arena WebDev, Text Arena Coding) with very different scores for the same model (e.g., Grok scores ~1450 on Text Arena Coding but ~1200 on Code Arena WebDev)
• Manual errors — we recently shipped gpt-5.2-codex at 1471 (the score for gpt-5.2-high, a different model) when its actual score is 1336

Proposed Solution

Agentically fetch ELO scores at generation time instead of maintaining a static dict:

1. Scrape or API-fetch the Arena leaderboard (e.g., arena.ai/leaderboard/code) to get current scores
2. Fuzzy-match Arena model names to litellm model IDs (e.g., gpt-5-medium → gpt-5, claude-opus-4-6 → claude-opus-4-6)
3. Fall back to the static dict only for models not on any leaderboard (local models like lm_studio/, niche providers)
4. Cache fetched scores with a TTL so we don't hit the leaderboard on every run

Context

The current ELO_SCORES dict has ~40 entries with inline comments like # [CODE] #22 and # [EST] to track provenance. Many [EST] entries are rough guesses. Fresh leaderboard data would eliminate all of these.

Acceptance Criteria

• generate_model_catalog.py fetches current ELO scores from at least one Arena leaderboard
• Graceful fallback to static scores if fetch fails (network issues, rate limits)
• Model name fuzzy matching handles Arena ↔ litellm naming differences
• Generated llm_model.csv has accurate, up-to-date scores without manual dict maintenance