[DRAFT] feat: add generalized Kaggle Community Benchmarks adapter

[mrshu]

Previously the only Kaggle ingestion was a single hardcoded script (utils/global-mmlu-lite) bound to one benchmark and one metric path. This commit adds a generalized adapter that converts any Kaggle Community Benchmark leaderboard into the EvalEval schema (v0.2.2), either for explicit owner/slug targets or by discovering and converting every published benchmark. A full run converts 996/1017 benchmarks into about 13.7k schema-valid records, which would make Kaggle the second-largest source in the datastore.

Discovery uses Kaggle's unauthenticated internal RPC (benchmarks.BenchmarkService/ListBenchmarks) with the anonymous XSRF cookie/header handshake, because no official package, sitemap, or public endpoint can enumerate community benchmarks. Per-benchmark leaderboards are read from the public REST endpoint
(/api/v1/benchmarks/{owner}/{slug}/leaderboard). Both APIs are undocumented for consumers, so the response shapes the adapter relies on are pinned by validation and tests rather than an upstream contract.

Schema mapping is deliberately conservative about what Kaggle actually exposes:

• Numeric results in [0, 1] become bounded continuous metrics; boolean results become binary. Numeric values outside [0, 1] have an unknown scale (Kaggle does not publish metric bounds) and are left untyped rather than given fabricated [0, 1] bounds; COUNTS-typed metrics are never bounded even when a value happens to fall in [0, 1].
• Confidence intervals are emitted as [score - ci, score + ci] because Kaggle's confidenceInterval is a symmetric half-width ("confidence radius"), matching the sibling utils/hle adapter.
• Each result carries its Kaggle evaluationDate as evaluation_timestamp when present (about 70% of results).
• On the discovery path, benchmark scoring config enriches each record: lower_is_better from sortOrder, metric_kind from aggregationType (only PERCENTAGE_PASSED -> pass_rate; generic reductions like AVERAGE/SUM are intentionally not metric families), metric_unit from displayType, with benchmark_id/aggregation_type/display_type recorded in source_metadata.additional_details. The targeted --benchmark path has no access to this metadata, so those fields fall back to defaults there (a documented limitation).

Bulk runs are built to fail loudly rather than silently lose data: a fetch failure is distinguished from an empty leaderboard, discovery truncation and per-benchmark conversion errors are isolated and surfaced, and the process exits non-zero when any benchmark could not be fetched, converted, or discovered, so a partial run is never reported as a clean success.

• Add utils/kaggle/adapter.py with --benchmark owner/slug (repeatable), --all, --limit, and --output-dir; targets are de-duplicated by (owner, slug), preferring the metadata-carrying entry
• Resolve the leaderboard owner as the organization slug when org-owned, else the creating user's username; owner-qualify evaluation_id
• Write to the standard data/{benchmark}/{developer}/{model}/{uuid}.json layout via the shared helpers (make_model_info, make_source_metadata, save_evaluation_log)
• Add tests/test_kaggle_adapter.py: 31 network-free tests covering result-type handling, CI bracketing, metadata enrichment, discovery pagination/filtering, the failure-vs-empty contract, dedup, and the non-zero-exit path
• Document usage, the discovery/bounds behavior, and the targeted-path limitation in utils/README.md; retain the global-mmlu-lite adapter for backwards compatibility