Layout — bench/ is now split by experiment instead of one flat pile of 25 files:
- common/     membw, cpuhog, freq_probe, make-inmate-initramfs.sh (shared by all four)
- rt-latency/ experiment 1 — RROS / Xenomai / PREEMPT_RT probes, same cell, same dtb
- crosscell/  experiment 2 — ivshmem-v2 link, heartbeat, D1/D2 detect-isolate-recover
- rt-demo/    experiment 3 — oob GPIO loop under a fully loaded rich Linux
- integrated/ experiment 4 — the same loop, now driven and monitored across cells

New in integrated/:
- rt_shm.h      single source of truth for the telemetry/command layout; its first 24 B
                are byte-compatible with crosscell's struct hb, so ivshmem-health keeps
                detecting D2 with not a line changed
- rt_ctrl.c     inmate side: oob GPIO loop + per-cycle telemetry publish + command poll
                + fault injection; hot spin touches neither shared memory nor barriers
- rt_console.c  root side: dash/get/set/sweep/inject-hang/clear-hang, each waiting for
                cmd_ack within one telemetry snapshot rather than assuming write == sent
- inmate-init-full, scenario_full.sh (T0–T8, every beat asserted, non-zero exit on any
  failure), tests/test_rt_shm.c
Plus common/freq_probe.c and one README per experiment directory.

Corrections folded in:
- the stimulus is the DHO924S built-in generator G, not an external AFG; its 5.00 V /
  1.000 kHz power-on default is called out because pin31 is a 3.3 V logic input
- hypervisor-enable is detected by the root-cell row, not by the exit code of
  `jailhouse cell list` — that exits 0 even when nothing is enabled, so enable was being
  skipped silently and cell create then failed with a near-undiagnosable EINVAL
- make-inmate-initramfs.sh: INIT_SRC= for one-pass init injection (repacking outside
  fakeroot loses the mknod'd dev/console), dd and taskset in the default tool set, and a
  TARGET= default matching the new directory depth
- the plaintext sshpass invocation is gone from scenario2_faultdemo.sh

Conventions:
- READMEs describe operations only — what each file is, how to build, deploy, wire and
  run it. Measured values and result analysis go in the competition document, not here.
- build outputs stay out of the working tree: every build line writes to "$ART", which
  defaults to ../artifacts, a sibling of the repo