Store: refresh a plan or text dimension's last_seen at most every 6 hours, not every hour

[erikdarlingdata]

Refs #4503.

Why

On one production store, collect.query_plan_dim's last_seen touch — the watermark a hot dimension row's re-sighting refreshes on every collection cycle it is referenced in — took 894,656 non-HOT updates in 71 hours under the original 1-hour guard, about 4.87 million full-page images, roughly 42 KB of WAL per row. last_seen is indexed (the prune's range scan needs it), so this update can never go HOT regardless of fillfactor, and dropping or reshaping the index isn't an option because the prune's ORDER BY last_seen LIMIT depends on it. The guard already caps a continuously-seen row to 24 touches a day; widening it to 6 hours caps the same row to 4 — a further 6x cut in how often this cost can recur per row.

What changes

• PayloadDimensions.LastSeenRefreshGuardHours is a new named constant (6), and PayloadDimensions.LastSeenRefreshGuardInterval renders it into SQL once. UpsertSql's four INTERVAL '1 hour' literals — the WHERE NOT EXISTS pre-filter and the ON CONFLICT ... WHERE guard, for both the compressed-content branch and the shared text/plan-xml branch — now interpolate that one constant, so both dimension tables (query_plan_dim, query_text_dim) and both content shapes take the same width from the same place.
• Doc comments on the guard and on DarlingRetention.ComputeDimensionCutoff/ComputeMapCutoff's margin discussion now name the constant and the measured WAL evidence in place of the old "one hour" wording.
• The dimension prune's retention margin (ChunkIntervalDays + 1 = 2 days) already covers this guard and the separate Query Store liveness-touch guard (currently 12 hours) with room to spare; it did not need widening. A new test class, PayloadDimensionGuardMarginTests, pins that both guards fit inside the margin directly — against the constants and against ComputeDimensionCutoff's actual cutoff arithmetic at several plan-content-retention widths — rather than leaving the relationship to comments that could drift.
• Existing live tests that exercised the old 1-hour boundary (PayloadDimensionLiveTests) were updated to the new 6-hour boundary: a re-touch at +3h now asserts no refresh, and +7h asserts a refresh, replacing the old +30min/+2h cases.

Test plan

• PayloadDimensionTests — the SQL string pins for UpsertSql, updated to INTERVAL '6 hours'.
• PayloadDimensionLiveTests — the churn-guard live test and the WHERE NOT EXISTS pre-filter live test, both re-timed to the new boundary; both GREEN on this branch.
• PayloadDimensionGuardMarginTests (new) — pure pins that the dim upsert guard and the Query Store liveness touch guard both fit inside the dimension prune's trailing margin, and that a row stamped at the wider guard's age survives the retention cutoff at several plan-content-retention widths.
• PlanContentRetentionTests, DarlingRetentionTests, DocCommentHygieneTests — unaffected pins re-run to confirm no collateral break.
• RED on dev at runtime: the new live test run against dev's 1-hour guard fails the +3 h case; last_seen is refreshed there.
• Mutation: LastSeenRefreshGuardHours set back to 1 on this branch → the same case RED; restored → GREEN.
• Darling.Tests builds with 0 warnings.

CHANGELOG

SECTION: Changed
ENTRY:

• The store writes less WAL maintaining its query plan and text dimensions ([Store: refresh a plan or text dimension's last_seen at most every 6 hours, not every hour #4502]) - a plan or query text that keeps being seen now refreshes its last_seen at most every 6 hours instead of every hour, so there are up to 6× fewer of these non-HOT updates. Retention is unchanged: the dimension prune's 2-day margin still covers the guard.
  REF:
  [Store: refresh a plan or text dimension's last_seen at most every 6 hours, not every hour #4502]: Store: refresh a plan or text dimension's last_seen at most every 6 hours, not every hour #4502