Refactor: CLI, results storage, model updates, and code quality fixes

[g-despot]

Summary

Major refactoring of the benchmark framework covering model updates, code quality fixes, results infrastructure, CI automation, and documentation.

Model updates

• Added latest models: Claude Opus 4.6, Claude Sonnet 4.6, Claude Haiku 4.5, OpenAI o3, o4-mini
• Updated default model references (judge, AnthropicModel default) from Claude 3.7 Sonnet to Claude Sonnet 4.6
• Updated weaviate-client from 4.16.7 to 4.20.4 (pyproject.toml + Dockerfile)
• Dockerfile updated from Python 3.9 to 3.10 to match pyproject.toml requires-python

Bug fixes

• Fixed env var warning crash: was calling .split() on None values instead of iterating over variable names
• Fixed GeminiModel: generation_config was built but never passed to the API call; now uses client.models.generate_content() with config=
• Fixed batch_import canonical implementation: referenced undefined collection variable instead of products (2 places), added missing import os and Auth import
• Fixed task ID grouping in summary output: multi-word task IDs like zero_shot_basic_semantic_search were split incorrectly; now uses prefix-based parsing

New features

• CLI (weaviate_vibe_eval/cli.py): argparse-based entry point with subcommands: run, list, leaderboard, compare, trends, runs
• Results storage (weaviate_vibe_eval/utils/results_store.py): Store results in a remote Weaviate BenchmarkRun collection with --store-in-weaviate flag
• Leaderboard generation: Ranked model table from stored results
• Run comparison: Diff two runs showing regressions/improvements per model+task
• Trend tracking: weaviate-vibe-eval trends --model <id> shows pass rate over time with UP/DOWN indicators
• Repetitions: --repetitions N runs each model+task N times with aggregated pass rate and duration stats
• LLM judge failure analysis: When --use-judge is enabled, failed tasks get diagnosed with root cause, failure analysis, and suggested fix
• Retry with backoff: All LLM API calls now retry 3 times with exponential backoff (2s/4s/8s)

CI/CD

• Added .github/workflows/monthly-benchmark.yml: Monthly scheduled run (1st of each month) + manual workflow_dispatch with configurable models/tasks/repetitions
• Results committed back to repo and stored in Weaviate

Code cleanup

• Removed unused TaskType class, get_model_info() method, _is_api_based attribute, parallel_models parameter, ThreadPoolExecutor imports
• Removed unused model_params, inputs, packages parameters from generate_and_execute()
• Removed stale requirements.txt (pyproject.toml + uv.lock is the single source of truth)
• Added missing weaviate_vibe_eval/utils/__init__.py
• Added logging module usage throughout (replaces bare print() for internal messages)

Documentation

• README.md: Complete rewrite with model/task tables, CLI usage, results storage docs, CI section
• CLAUDE.md: New technical implementation guide covering architecture, file map, result formats, env vars, and docs dashboard integration
• JSON output now includes metadata envelope (run_id, git_commit, timestamp, model list, repetitions)
• Markdown reports include summary table with pass rate and duration columns