runner: compare Criterion SQL-harness benchmarks with critcmp

[adriangb]

Problem

bench.sh has two benchmark families, and the standard runner only knew how to compare one:

Family	Output	Compared by
dfbench-based (clickbench, imdb, h2o, tpcds, sort_*)	results/<name>/*.json	bench.sh compare_detail ✅
Criterion SQL harness (wide_schema, predicate_eval, run via cargo bench --bench sql)	target/criterion/**	nothing ❌

The standard runner always called compare_detail, which reads results/*.json. The SQL-harness suites never write those, so the comparison was empty — the result comment showed:

Comparing HEAD and <branch>
Note: Skipping .../results/HEAD/*.json as .../results/<branch>/*.json does not exist

i.e. only resource usage was posted, no per-query timings — confirmed on a real wide_schema run.

Fix

Route the SQL-harness suites through the same mechanism bench_criterion.rs already uses for native Criterion benches:

1. Run each side with SQL_CARGO_COMMAND=cargo bench --bench sql -- --save-baseline <RESULTS_NAME> so Criterion saves a named baseline per side (HEAD for base, the branch name for the PR). The SQL harness's arg parser uses ignore_errors, so --save-baseline passes through to Criterion.
2. After both sides, gather the baselines into one target/criterion tree (reusing bench_criterion::copy_criterion_baselines) and diff them with critcmp HEAD <branch>.

dfbench suites are unchanged (still compare_detail). A run mixing both families emits both sections. is_criterion_harness() enumerates the SQL-harness suites and generalizes to future ones.

No datafusion-side changes and no new dependencies — critcmp is already in the runner image and used by the criterion path.

Testing

cargo fmt --check, cargo clippy --all-targets -- -D warnings, and cargo test (135 tests, incl. a new criterion_harness_classification) all pass. End-to-end verification will come from a wide_schema run once deployed (the suite already runs to completion; this is purely about producing the comparison output).

🤖 Generated with Claude Code

[adriangb]

It'd be nice if these weren't hardcoded...