CF 7392, v7: benchmarks for 0005 and 0006
==========================================

Host: c8id.4xlarge (Xeon 6975P-C), Debian 13, gcc 14.2, -O2, no cassert,
everything on local NVMe.  Builds and binary hashes are in summary.txt.

  p4    82a84654f12  v7 0001-0004
  base  05b22c4ef16  v7 0001-0005
  new   base + 0006 (identical to v7-0006 except it was a working-tree
        diff when measured)

Execution: bench5.sh.  One data image (b5-data.sql) copied fresh for each
run so both builds read the same files; every relation pg_prewarm'd;
AB/BA order randomized per block from SEED; one discarded run, then 5
timed, per query per variant.  an5.py takes each block's median per
variant, the log ratio, a t interval over blocks and a sign-flip p.
Two sessions of 11 blocks each.

  g1.csv, g2.csv   base vs new  (0006's effect), queries in b5.q
  h05a.csv, h05b.csv   p4 vs base (0005's effect on Heikki's shape),
                       queries in b5h.q; summary and Heikki's exact
                       10k-row example (heikki-exact.sh) in
                       heikki-0004-vs-0005.txt

The function in all of them is fexp(), plpgsql that loops 50 times and
is declared COST 10000.  The planner settings are default except
max_parallel_workers_per_gather = 0.  With parallel query allowed, 0004
already finds a parallel index-only scan for the 100k-row ORDER BY case
(8.6 ms vs 7.5 ms with 0005).

Planning time: plantime2.sh, N-way joins (atab and dim alternating),
median of 50 EXPLAINs per query, base vs new, three target lists: a
plain column, fexp(t0.i), and abs(t0.i) + 1 with an index on it (cheap,
under the cost gate).  Output in plantime-v6-vs-0006.txt.

Join sweep: joinsweep.sh, the same N-way joins with fexp(t0.i), median
of 5 runs, execution and planning ms, base vs new; output in
joinsweep-0005-vs-0006.txt.

Correctness: correct5.sql, 35 query shapes (joins of every kind,
outer-join nullable sides, partitioned with differing column order,
partitionwise, lateral, CTE, UNION, parallel-unsafe function...), each
run with index-only scans on, off, parallel encouraged, partitionwise
joins on, nested loop only and merge join only; every result is
compared with the off run.  0 mismatches in 210 runs, cassert build.
