Version n20260805

Ratings process rewrite, security upgrades, backend parity

Deployed to GAE as version n20260805.

/stats/ratings rewrite (#149)

The nightly top-100 ratings computation no longer scans the entire StatsModel history (3.5M entities, ~11 minutes) with false positives that could truncate the lists:

• Exact current tables from indexed UserModel Elo fields plus parallel newest-snapshot lookups (StatsModel.newest_before_multi) — false positives are structurally impossible
• Daily archives — each night's tables stored as JSON in the new RatingArchiveModel kind (PG: rating_archive); yesterday/week/month deltas become keyed lookups, no new composite indexes
• /stats/ratings_backfill — self-chaining Cloud Tasks endpoint that fills historical archives one date per task, idempotent, pinned to the executing service/version so it can run on a not-yet-promoted deployment
• ~30–50× faster overall (measured ~15× on explo-dev; netskrafl runs should drop from 11 min to well under 30 s)
• NDB/PG parity fixes: kind-specific fields in PG _list_by_elo, robot stats in PG newest_before, module-level put_multi in the stats job, manual memcache invalidation
• 24 new backend-parity tests

Dependency security upgrades

Flask, Authlib and requests upgraded, resolving 18 security advisories.

NDB↔PostgreSQL backend parity (#148)

Strengthened parity between the NDB and PostgreSQL backends with expanded tests, plus a duplicate-email diagnostic utility.

Operations

• explo-live scaling optimized; basic-scaling comment fixed in app-netskrafl.yaml
• Deploy skill with gradual (canary) traffic rollout and log monitoring