Incorrect avg aggregation result when querying CSUP data (regression at 66b20d0)

[philrz]

tl;dr

Running ClickBench query 2 against CSUP data produces an incorrect avg() aggregation result. The incorrect result comes in both the sequential runtime and vector runtime, so it's presumably a problem with the conversion to the CSUP format. I've confirmed that this problem showed up starting at commit 66b20d0, which is associated with the changes in #5577.

Details

Repro is with super commit 04b7efc using the hits.parquet test data from ClickBench.

As a baseline, here's the presumed correct avg() result of 1513.4879349030107 that matches when querying the original Parquet data from DuckDB and SuperDB using both the sequential and vector runtime.

$ duckdb -version
v1.1.3 19864453f7

$ duckdb -c "SELECT SUM(AdvEngineID), COUNT(*), AVG(ResolutionWidth) FROM hits.parquet;"
┌──────────────────┬──────────────┬──────────────────────┐
│ sum(AdvEngineID) │ count_star() │ avg(ResolutionWidth) │
│      int128      │    int64     │        double        │
├──────────────────┼──────────────┼──────────────────────┤
│          7280088 │     99997497 │   1513.4879349030107 │
└──────────────────┴──────────────┴──────────────────────┘

$ super -version
Version: v1.18.0-237-g04b7efce

$ SUPER_VAM=1 super -c "SELECT SUM(AdvEngineID), COUNT(*), AVG(ResolutionWidth) FROM hits.parquet;"
{sum:7280088,count:99997497(uint64),avg:1513.4879349030107}

$ super -c "SELECT SUM(AdvEngineID), COUNT(*), AVG(ResolutionWidth) FROM hits.parquet;"
{sum:7280088,count:99997497(uint64),avg:1513.4879349030107}

Once we convert the Parquet file to CSUP, the avg result is now very different in both the sequential and vector runtimes.

$ super -f csup -o hits.csup hits.parquet

$ SUPER_VAM=1 super -c "SELECT SUM(AdvEngineID), COUNT(*), AVG(ResolutionWidth) FROM hits.csup;"
{sum:7280088,count:99997497(uint64),avg:7224.820559888614}

$ super -c "SELECT SUM(AdvEngineID), COUNT(*), AVG(ResolutionWidth) FROM hits.csup;"
{sum:7280088,count:99997497(uint64),avg:7224.820559888614}

Tests indicate this started happening at commit 66b20d0, which is associated with the changes in #5577. The result was correct at the commit just before:

$ super -version && super -f csup -o hits.csup hits.parquet && SUPER_VAM=1 super -c "SELECT SUM(AdvEngineID), COUNT(*), AVG(ResolutionWidth) FROM hits.csup;"
Version: v1.18.0-226-g02d23319
{sum:7280088,count:99997497(uint64),avg:1513.4879349030107}

Then incorrect at 66b20d0.

$ super -version && super -f csup -o hits.csup hits.parquet && SUPER_VAM=1 super -c "SELECT SUM(AdvEngineID), COUNT(*), AVG(ResolutionWidth) FROM hits.csup;"
Version: v1.18.0-227-g66b20d0f
{sum:7280088,count:99997497(uint64),avg:7224.820559888614}