release: return processed grains from rollups

[NikolayS]

Fixes #191.

Unblocks #192.

Contract decision

ash.rollup_minute(integer) and ash.rollup_hour() now return the number of completed time grains processed, independent of how many (grain, datid) rows are upserted. This matches their catalog comments, the batch-limit model, and the operator-visible meaning of catch-up progress.

Each completed loop iteration counts once, including an empty grain: it consumes one batch slot and advances its watermark. A call with an already-caught-up watermark still returns exactly 0 because it processes no loop iterations.

Changes

• increment rollup results once per processed minute/hour rather than by DML ROW_COUNT
• add exact CI assertions for one and two databases, physical row counts, watermark deltas, empty grains, and already-caught-up rollup_minute(1000000)
• align the historical rollup blueprint and 1.5-to-2.0 upgrade comment with the corrected contract
• disclose the user-visible return correction in RELEASE_NOTES.md

Call-site audit

• ash.rotate() stores the minute-rollup result but never reads or compares it. Exceptions and its independent (minute, datid) completeness query remain authoritative, so release: prevent cleanup from deadlocking rotation #185's truncation safety is unchanged.
• ash.rollup_cleanup() does not consume either return value; it only shares the rollup advisory lock and continues returning physical deletion counts.
• ash.status() does not call either rollup function; it reports physical table rows and watermarks directly.
• pg_cron and documented external scheduler bodies execute bare SELECT statements and do not branch on the scalar.
• existing tests either discard results, assert 0 on no-work/busy/incomplete paths, assert > 0, or assert 1 for a single grain/database. Those expectations remain valid. PR release: add behavioral feature-gate coverage #192's retained two-database assertion changes from RED to GREEN.

Validation

After the required hand rebase onto main 3af54a7:

• exact issue rollup_minute() and rollup_hour() return per-database row counts, not processed grains #191 RED on main: two datids produced two physical rows and returned 2 for one minute/hour grain
• exact issue rollup_minute() and rollup_hour() return per-database row counts, not processed grains #191 GREEN on this head: two physical rows remain, both functions return 1, and watermarks advance by 60/3600 seconds
• exact added contract block: PASS for one/two datids, empty grains, and already-caught-up 0
• complete release: prevent cleanup from deadlocking rotation #185/Release blocker: rollup_cleanup can permanently deadlock rotation #162 cleanup and rotation regression: PASS
• independent static caller/contract review: no blocker
• composed with release: add behavioral feature-gate coverage #192: all four feature-bearing surfaces PASS on both PG17 and PG14 (8/8)
• workflow YAML parse and git diff --check: PASS

The migration header was also corrected during rebase so it no longer broadly claims all admin/lifecycle functions are unchanged while documenting this intentional return-value correction.

[samo-agent]

Verified on exact rebased head 717545e. Issue #191 is reproducibly RED on main (one processed grain with two datids returns 2) and GREEN here (returns 1 while preserving two physical rows and 60/3600-second watermark movement). The exact added contract covers one/two datids, empty grains, and caught-up zero-work. The complete #185/#162 rotation regression passes. Independent static review found no blocker, and composing this fix with #192 makes all four feature-bearing surfaces pass on both PG17 and PG14 (8/8). Migration wording was corrected to describe the intentional return-value change without claiming all lifecycle behavior is unchanged. APPROVED.