Benchmark docs drift found by the weekly audit of 2026-11-11: four decisions that are not text fixes #589

The weekly audit of 2026-11-11 compared the benchmark docs on main (b921765) with the harness code. The pull request linked below fixes the text drift: three wrong statements (which client is the served headline, two function names that do not exist), references to code that #432 removed, the CrateDB memory rule, the status of a few section 7 rows, and the statement about when the mini pulls main. What is left are decisions, not text fixes:

• Idle-exit latency: the docs say the runner holds it at 10 microseconds, the tree default is 0 and only the stage scripts set 10. Either the docs say so, or 10 becomes the tree default, which is a mid-run change because the comparator stages pull main at each start. My default is to follow the code in the docs.
• Row 93 (the wheels-only re-measure of the ingest cells), CAMPAIGN section 7: add the row for the wheels-only post1 re-measure of the ingest paths #581, docs(campaign): row 93, the wheels-only post1 re-measure of the ingest paths #582, benchmarks: tooling for the post1 rerun of the paper's embedded cells (build-only mode, wheel label, one wheel per cell, embedded-only stages) #583 and benchmarks: tooling for the post1 rerun of the paper's embedded cells (build-only mode, wheel label, one wheel per cell, embedded-only stages) #584 still say post1. The wheels that exist are post2, published on 2026-11-10. Retarget them.
• Row 88: run the PQ arms in the 2026-11 pass now that 26.11.1 carries the fix, or not.
• Rows 6, 11 and 17: stay on the 2026-11 list, or move to the next campaign.
• Row 19: close it as superseded by Next instrument: memory and disk per phase, a JVM working set after GC, exact peaks (for the first full run after the paper) #352, if the page prints disk_<phase>_data_mib.