The partition declaration’s cost claim has no fixture evidence left

Decide, on evidence, whether the partition dimension’s declared selectivity still earns its place, and leave whichever answer is true holding a measurement rather than a memory. The declaration is a statement the model makes about one column’s distribution so a pass planning inside a transaction, where no ANALYZE can run, reaches an index it would otherwise price badly. Its cost justification used to be a control in PartitionSelectivityWorthTest: the same read visited 10939 rows without the declaration against 1113 with it. Three conversions have since put primary keys on the relations that reader joins, each key giving the planner a seekable ordering it can reach without being told anything, and the gap closed to twelve rows in 524. The control was retired rather than propped up with a lower floor, so the file now records a collapse and the declaration’s remaining evidence is plan text in RefreshPlanStatisticsTest plus one consumer-scale round taken on a shape that no longer ships.

That is a real gap and it points both ways, which is why this is a question and not a repair. The fixture understates by construction: a per-driving-row partition scan is linear in driving rows and a twelve-unit fixture is precisely where such a scan is cheap, so the declaration may well still be worth a great deal at consumer scale with nothing here able to see it. Equally it may have been made redundant by the keys, in which case a declaration nobody can measure is complexity the model is carrying for a defect it already fixed another way. Landing this means either an instrument that has signal again, on a fixture sized to where the cliff lives or on a reader whose keys do not reach it, or a decision to retire the declaration with the reasoning written down. What must not survive is the current state, where the claim stands on a measurement whose subject has moved.