Archived leaderboard pagination returns 429 during benchmark collection

[Andyyyy64]

During benchmark collection with 0.5.18, the archived Open LLM Leaderboard v2 endpoint returned HTTP 429 at offset 3600, with pages of 100 rows. Collection completed with 263 scores from the available sources.

Please investigate the request volume for this frozen dataset and the handling of server rate-limit responses. The scope is this source's fetch behavior, not a new mirror system or a rewrite of ranking.

This is a concrete follow-up to #83. It is not established as a regression introduced by 0.5.18.

[Andyyyy64]

The retry and partial-page fix from #182 is included in v0.5.20 (https://github.com/Andyyyy64/whichllm/releases/tag/v0.5.20). If a later leaderboard page is still rate-limited after bounded retries, the source retains the pages already fetched. A first-page failure still fails the source.

The warning goes to stderr. Partial results use the existing 24-hour cache, and a cached read does not repeat the fetch warning. Use --refresh to fetch the scores again. The public 0.5.20 package completed an empty-cache recommendation after a 429 at offset 3600, retaining 3539 scores. Its JSON parsed, and a cached rerun returned the same picks without repeating the warning.