[tamnd]

Part of #1.

What changes

• src/instructions.rs is the instruction count runner of spec/21 §21.14. It runs each query of a set under perf stat -e instructions -x ,, the counter that the spec names. Each query runs 3 times and the runner keeps the smallest count.
• For DuckDB, perf counts the duckdb process, with SET threads = 1. For a server, perf counts the cgroup of the server unit on all CPUs with -G, and the client is psql. Background processes of the server add to that count, so the runner also counts SELECT 1 3 times and subtracts the smallest of those counts from each query.
• rupg-bench instructions --queries PATH reads the set from a file with one query on each line, as in ClickBench, or from a directory of .sql files, as for TPC-H. --set names the set. --engine names the engine and is part of the suite name, for example instructions-clickbench-postgresql.
• --ratchet ratchet.toml compares the counts with the table of that engine and set. The command fails when one query is more than 3 percent over its row or the total is more than 1 percent over. --save writes the counts as the new table.
• ratchet.toml is new. It has no table yet. The counts must come from a dedicated runner (spec/21 §21.14), and server3 is a shared machine.
• reports/2026-10-07/6005c2b1-server3-instructions-clickbench-{postgresql,duckdb}-smoke.{md,json} are the reports of the two smoke runs below. Their names and their first line mark them as smoke runs.
• The README explains the command.

Test

srvtest passed on server3: clippy is clean and 43 tests pass. The tests cover the parser of the perf output, the query files and the budgets.

Smoke runs on server3, 7 October 2026. These are not baselines and not the ratchet. server3 is a shared machine with 8 cores, and its load average was 66 to 85 that day. The data is the first 1,000,000 rows of ClickBench (hits_0.parquet). DuckDB 1.5.6 read it into hits1m.db, and PostgreSQL REL_19_STABLE at 7d3d2db7 read it with \copy into the table hits of the database bench. The query files are queries.sql of the ClickBench pin e6bda4e for each system.

PostgreSQL, 10:05:06 to 10:32:42 UTC:

RUPG_BENCH_COMMIT=6005c2b1 rupg-bench instructions --set clickbench --queries pg-queries.sql --engine postgresql --unit postgresql@19-main --smoke --report /root/work/reports --machine server3

DuckDB, 10:32:42 to 10:46:12 UTC:

RUPG_BENCH_COMMIT=6005c2b1 rupg-bench instructions --set clickbench --queries duckdb-queries.sql --duckdb hits1m.db --duckdb-bin /root/work/rupg-bench-opt/bin/duckdb --smoke --report /root/work/reports --machine server3

Number	PostgreSQL	DuckDB
base, SELECT 1, smallest of 3	12,103,536	78,040,265
total of the 43 queries, net	228,920,014,998	18,064,983,237
largest query	q28, 54,758,339,550	q28, 8,580,483,958
median spread of the 3 runs of a query	0.62 percent	4.45 percent
largest spread	q0, 77 percent	q6, 63 percent

The spread is (largest run minus smallest run) divided by the smallest run. On this shared machine the DuckDB spread is above the 3 percent budget. That is why the ratchet table stays empty until a dedicated runner exists.

The instructions line of #1 stays open. The fixed set also needs the 22 TPC-H queries at SF0.1, 1,000 YCSB point reads and 100 TPC-C new-order transactions, and those drivers come in the next PRs.