Memory Leak on ChunkReader.newBuf

We have been seeing a memory leak that shows up over time, that we can't seem to get to the bottom of, specifically when in a transaction, and we do a SELECT during that query. This process is being run in parallel ~100 at a time, and will develop a significant leak after 5-10 minutes of running, that never gets cleared up.

The code that primarily shows the memory being retained (when reallocating the chunk reader buffer location). https://github.com/jackc/chunkreader/blob/3c4a99247c6656284dd2be8d74438ca9bbb4f342/chunkreader.go#L72-L76

pgx portion of the trace:

Any insights that could help us narrow down what may be retaining the reference?

Start with chunkreader.go lines 72-76 and the linked heap trace, then reproduce the reported workload: roughly 100 parallel processes performing SELECTs inside transactions for 5–10 minutes. Trace the retained buffer reference and document the cause or a focused regression test demonstrating the leak.

Written by the indexing model from the issue text.