[linuxnow]

Closes #188. Part of #182. Also covers the xrun and latency fields asked for in #192.

Base: feat/bench-harness (PR #179 has not merged). Retarget this PR to main when #179 lands.

The grouping design note (docs/design/notes/2026-10-27-grouping-planner.md, "The benchmark") asks for the bench to run a strip as individual, chain, plan and pool, to report xruns and a stress profile, and to repeat each row three times with the median and the spread. This adds that, and a script that makes the grouped-vs-processes table from the documents.

What changes, in small commits:

• Schema (still omx-bench/1, fields only): xruns, latency_frames, stress, repeats, spread, and the modes plan and pool. A spread has to hold its median, a stressed row has to report xruns, and the same scenario with and without stress is two rows.
• xruns and latency_frames: a client on the run's private JACK server counts xrun callbacks over the measurement window (libjack through ctypes, no port, no process callback) and reads the capture latency at the playback port. If the server's own log names xruns the probe never heard, the row fails instead of reading 0.
• Stress: one busy loop pinned to each core the host may use, plus a process that touches 256 MiB, for the whole window. The row carries busy<cores>+mem256. --stress none|on|both, both by default for the cvp and grp scenarios.
• Repeats: --repeats (3 by default). Median per measure, a count takes the higher middle run, [min, max] under spread. A failed repeat fails the row and says which.
• plan and pool: new family grouping-vs-processes (grp8, grp16, CLAP plugins, omx-drive pinned isolated). The plan and pool rows run with pins on and a verdict per plugin written into the daemon's own pin store, grouping on, and the package pins off. A daemon without the grouping setting gives an unsupported row; a plan of one segment, or a pool the daemon did not form from the members moved into it, gives a failed row.
• bench/table.sh: figures with their spread, the individual-over-chain ratios, and xruns summed per mode with and without stress.

Tests (python3 -I bench/tests/test_bench.py, 51 tests, no JACK and no host needed) and the sabotage of each guard in bench/tests/sabotage.sh (34 new cases, each breaks one line in a copy and requires its test to go red). CI runs both.

Not done here, on purpose:

• Nothing was run against JACK or the Pi. The ctypes calls to libjack are tested against a stand-in, and the plan and pool steps against a stand-in daemon, so the first real run on a desk or the Pi is the check of those two seams.
• The plan, plan_get verbs and the grouping setting are those of the design note; the planner (L2 grouping: the planner and the plan, plan_get, plan_clear verbs #185) and the worker's xruns line (omx-clap-core#371) are not merged, so on today's daemon the plan and pool rows are unsupported.
• The cross-chain pool is not measured: one strip is one chain, so the pool holds that strip's leftovers.