fix: do not divide parallel custom-scan I/O by worker count, with the leader arm covered (#1065, rebased + coverage)

[OffgridwithJD]

This is @linuxhikerpm's #1065, rebased onto cfdb393 with the test-coverage gap closed and its ledger re-derived. Their commits are preserved; mine is the last one. Opened from my fork because I cannot push to theirs.

The C is legitimate and correct

pgcolumnar_parallel_divisor reproduces core's get_parallel_divisor (costsize.c) faithfully, and separating I/O from CPU is the right fix — core's seqscan divides CPU and leaves the disk run whole. The path->rows change is correct and belongs with it. Nothing to argue with in the code.

The defect was in the tests, and it hid the main branch of the copied heuristic

Both suites set parallel_leader_participation = off, and that GUC defaults ON. With it off the divisor is exactly the worker count, so this arm —

if (parallel_leader_participation) { leader_contribution = ...; parallel_divisor += ...; }

— the entire reason the function is not a one-line division, never executed in either harness. The copied heuristic was covered only in the configuration nobody runs and untested in the one everybody does. Same shape as setting-a-guc-is-not-engaging-the-path: the setting is present, so the coverage looks real, and the branch is dark.

What I added

Both suites now measure the same property a second time with the leader participating, and prove the branch ran rather than asserting it did. The row estimate is the observable: the partial path divides rel->rows by the same divisor, so leader-on and leader-off cannot agree.

Measured on PG17, two workers, 40,000 rows:

rows:   leader off = 20000     leader on = 16667      (40000/2 against 40000/2.4)
ratio:  leader off = 1.070     leader on = 1.083      (bound is 1.35)

If the leader-participation branch changes the divisor ever reports same, the branch did not run and every assertion after it is about the wrong divisor.

The suites stay independent: the shell arm reads the plan text, the pytest twin walks FORMAT JSON, neither invokes the other, and the names match so compare_to_bash grades them one-for-one.

Verified

parallel_scan_cost.sh          13/13 on PG15/16/17/18/19
test_parallel_scan_cost.py     13/13 against a real cluster on PG17
compare_to_bash                128/128 one-for-one
shellcheck -S error            clean
ledger 1243 rows uniform 15;16;17;18;19, census 1235, GATE rc=0

Ledger rows re-derived from those runs, not by editing field 4. Their rows read 15;16;17;18 while every other row reads 15;16;17;18;19, which the release gate refuses even though CI is green (ci.yml:503).

Merge order

Takes TESTS.md section 44, correct while main is at 43. All five take 44; the numbering is gated, so the second to land renumbers and re-derives its census — the conflict offers two numbers and neither is right (#996).

Suggested order: #1092 (their #1063) first, then #1094/#1095, then this, then #1093.

Original: #1065. Author: @linuxhikerpm.

[jdatcmd]

Approving.

The divisor is a faithful copy and I checked it against core rather than from memory. Diffed against get_parallel_divisor in src/backend/optimizer/path/costsize.c of a real PostgreSQL source tree: same initialisation, same parallel_leader_participation guard, same 1.0 - (0.3 * parallel_workers), same > 0 test, same accumulation. Logic-identical.

Separating I/O from CPU is right, because core's seqscan divides CPU and leaves the disk run whole.

The coverage fix closes a real gap. On #1065 both harnesses set parallel_leader_participation off and nothing set it back, so the if (parallel_leader_participation) arm at columnar_customscan.c:2531 never executed. The GUC defaults ON, so the only covered configuration was the one nobody runs.

The new arm is a genuine differential rather than a GUC flip: check "the leader-participation branch changes the divisor" asserts rows_off != rows_on, so it fails if the branch is inert. That is the part that makes the coverage real.

Commit preservation checked: every commit from #1065 is an ancestor of this branch.