Document stateless popular sweep behavior

crates/osu_fetcher/README.md

@@ -73,7 +73,7 @@ editing the file.
 
 | Command | What it does |
 |---|---|
-| `enumerate` | Paginates `/api/v2/beatmapsets/search?m={mode}&s={status}&sort=updated_desc` for each requested status, upserts every beatmapset into `sets`. Cursor checkpoints saved per-status after each successful page → can resume mid-run. **Incremental by default**: each run saves a `last_updated` high-water mark and the next run stops paginating once it crosses that mark; pass `--full-rescan` to re-walk everything. For popular maps, use `--statuses any --sort favourites-desc --min-favourites 100 --min-playcount 1000 --stateless`; that walks the favourite-sorted prefix, filters low-play rows, and stops below the favourite threshold. |
+| `enumerate` | Paginates `/api/v2/beatmapsets/search?m={mode}&s={status}&sort=updated_desc` for each requested status, upserts every beatmapset into `sets`. Cursor checkpoints saved per-status after each successful page → can resume mid-run. **Incremental by default**: each run saves a `last_updated` high-water mark and the next run stops paginating once it crosses that mark; pass `--full-rescan` to re-walk everything. For popular maps, use `--statuses any --sort favourites-desc --min-favourites 100 --min-playcount 1000 --stateless`; that walks the favourite-sorted prefix, filters low-play rows, persists matching set IDs as normal wanted rows, and stops below the favourite threshold. Because it is stateless, repeating it re-queries the popular prefix but does not re-download already-successful rows. |
 | `discover` | Walks mirror search/list surfaces for ranked/approved/loved/qualified candidates and promotes sets to `wanted` when at least `--min-quorum` independent mirrors agree. Use after `enumerate` to catch cached sets not surfaced by official API search. |
 | `scan` | Walks `--archives-dir`, parses `<set_id>.osz` filenames, marks each as `success` in the state DB. Optional `--verify` opens each archive and walks its zip entries; corrupt files stay pending. |
 | `download` | Runs a bounded greedy worker pool. Each idle worker claims the next `pending wanted` row in `set_id` order and cycles mirrors per-set. See **Failure modes** below for the resilience guarantees. |

docs/PIPELINE_RUNBOOK.md

@@ -117,8 +117,20 @@ sorts by `favourites_desc` and stops when the page stream falls below
 `POPULAR_MIN_FAVOURITES`, while also filtering out rows below
 `POPULAR_MIN_PLAYCOUNT`. This is intentionally more expensive than the normal
 updated-desc enumerate, but it walks only the popular prefix rather than the
-whole search catalog. Set `POPULAR=0` to disable it for a quick maintenance
-run.
+whole search catalog.
+
+The popular sweep persists its results as normal wanted rows in the fetcher
+state DB, so already-acquired popular sets are not downloaded again. The sweep
+itself is still stateless: it does not store a popular-sweep cursor, high-water
+mark, or "completed recently" checkpoint. Each normal update reruns the
+favourite-sorted API walk and idempotently upserts rows it has already seen.
+
+Set `POPULAR=0` for retry-only or quick maintenance runs, especially when the
+goal is just to retry failed downloads already present in `state/fetcher`.
+
+```bash
+POPULAR=0 DOWNLOAD_CONCURRENCY=4 bash scripts/update_maps_v1.sh
+```
 
 ## Metadata Compaction
 
@@ -268,6 +280,10 @@ drains. That second pass is skipped when `DOWNLOAD_LIMIT` is set so smoke-test
 limits remain meaningful. Rows that all mirrors reported absent are left alone
 unless you set `RETRY_MISSING=1`.
 
+For a failed-download retry run where you do not want to re-walk the popular
+favourites prefix, set `POPULAR=0`. Existing wanted rows and failed rows remain
+in the hydrated fetcher state DB.
+
 If ingest is interrupted, rerun the same script with a fresh `BATCH_ID`.
 Committed archives are skipped by SHA-256. An incomplete chunk has no
 `archive_revisions` commit marker and is retried.

docs/osu_fetcher_guide.md

@@ -25,10 +25,22 @@ The script:
 
 - hydrates compact `data/`, `schemas/`, and fetcher state from the bucket;
 - seeds existing compact beatmapset IDs as already acquired;
-- runs osu! API enumeration and a small ranked-front scan;
+- runs osu! API enumeration, a small ranked-front scan, and the popular
+  favourites sweep unless `POPULAR=0`;
 - downloads only pending new or changed sets to `incoming_osz/`;
 - ingests the archives, compacts metadata by default, and publishes the update.
 
+The popular sweep stores matching set IDs as normal wanted rows in
+`.fetcher/state.db`, but the sweep is stateless. It does not store a cursor or
+"already completed" marker, because favourite/play counts can change without a
+beatmapset `last_updated` change. Normal updates rerun the favourite-sorted
+prefix and upsert already-known rows idempotently. For retry-only runs, skip
+that API walk:
+
+```bash
+POPULAR=0 DOWNLOAD_CONCURRENCY=4 bash scripts/update_maps_v1.sh
+```
+
 Archive publication uses `hf buckets sync archives ... --ignore-existing
 --no-delete`. The archive prefix is content-addressed and append-only, so a
 cancelled upload can be resumed by rerunning the wrapper with `HYDRATE=0` when