fix(dq): partition decoded tables by real dates, not day-of-month

[elffjs]

Problem

day(x) in DuckLake is a component extraction — day-of-month 1–31, not a date truncation — so PARTITIONED BY (subject_bucket, day("timestamp")) on lake.signals / lake.events produced 31 cycling time buckets that mix every month forever (day=9 collects Jul 9, Aug 9, Dec 9, …). Partitions never close, and the anti-join dedup probe's "day partition" pruning was actually pruning on day-of-month.

Fix

Use the documented DuckLake daily-partition shape — the year/month/day triple:

ALTER TABLE lake.signals SET PARTITIONED BY (subject_bucket, year("timestamp"), month("timestamp"), day("timestamp"))
ALTER TABLE lake.events  SET PARTITIONED BY (subject_bucket, year("timestamp"), month("timestamp"), day("timestamp"))

The signals_latest / events_latest rollups stay subject_bucket-only, unchanged.

Notes

• Known trade-off (documented in setupStatements): ducklake_merge_adjacent_files only consolidates within one partition, so a low-traffic (bucket, day) keeps its tiny files forever. Accepted for layout legibility.
• Deploy: the ALTERs only run at table creation, so an existing catalog keeps the old spec — deployed environments (all droppable per team) should start from a fresh catalog.
• Companion PR: fix(din): partition raw_events by real dates, not day-of-month din#10 (same bug on raw_events).

Testing

Full go test ./... passes.

🤖 Generated with Claude Code