Document stall and rollback handling

README.md

@@ -61,6 +61,7 @@ Built from the daily snapshots of [cfahlgren1/hub-stats](https://huggingface.co/
 - `dl30` is the Hub's rolling 30-day download count. `dl_all` (all-time downloads) only exists from 2025-02-27, so exact daily downloads, computed as the difference of `dl_all` between snapshots, start on that date.
 - Some days are missing in the source (Aug 2024, Jun 2025, Apr 2026, May–Jun 2026). Totals are unaffected; daily values across a gap are averages.
 - A model or dataset is tracked once it has 10+ downloads in 30 days, 50+ all-time downloads, or at least one like. A Space is tracked once it has a like; the Hub doesn't publish visits for Spaces.
-- On some days the Hub's download counters stand still and catch up a day or two later. `hub_series` spreads each such episode evenly over its days (sums unchanged), and the snapshots inside it are listed in `skip_days` (`meta.json` for models, `repos_meta.json` for datasets) so per-repo series can do the same.
+- On some days the Hub's download counters stand still and catch up a day or two later, mostly on Wednesdays. Dataset counters freeze all at once; model counters often freeze only in part (newer repos stop, older ones keep counting), so the model total may only dip to half. A day is treated as a stall when the total is under 0.3 of its 15-day median, when at least 25% of the models with 60K+ monthly downloads didn't move at all, or when a dip under 0.7 is made up by the next or previous day. `hub_series` spreads each episode evenly over its days (sums unchanged), and the snapshots inside it are listed in `skip_days` (`meta.json` for models, `repos_meta.json` for datasets) so per-repo series can do the same. `frozen` in the same files holds the recent daily share of frozen counters.
+- Twice (May 2025, June 2026) the all-time counters went down for a large share of models and came back days later. Those snapshots are in `skip_days` too, and downloads across them are measured from the last good snapshot to the first one after the counters recovered, so the recovery isn't counted as new downloads. A drop that doesn't recover within 8 snapshots is treated as a lasting correction by the Hub, not set aside.
 - `uses.parquet` comes from the cards as they are today, dated by when each Space or model was created, not by when it started using what it lists.
 - Download counts follow the Hub's [counting rules](https://huggingface.co/docs/hub/models-download-stats); the Hub occasionally books delayed downloads on a single day.