Context

While experimenting with ways to reduce noise in the PR benchmark data, Claude found a number of bugs in the existing benchmark code:

• benchmarkSingleInsert always uses the smallData file instead of the source param, meaning BenchmarkLargeDocInsertOne and BenchmarkSmallDocInsertOne use the same fixture file.
• benchmarkSingleInsert doesn't call ResetTimer, so the benchmark erroneously measures the test setup time as well as the benchmark loop.
• testdataDir has a buggy file traversal termination condition and is generally unnecessary. It can be replaced by a path constant.
• setupBench never calls the CancelFunc for the Context used to Ping.
• The "benchmark" task in the Taskfile references a package that no longer exists.
• The test fixtures are only downloaded if the benchmarks are invoked via TestRunAllBenchmarks. If the benchmarks are run any other way, the fixtures are not downloaded.

Definition of done

• Fix all bugs listed above.

Pitfalls

What should the implementer watch out for? What are the risks?