[bompus]

Adds skills/hr-benchmarking/scripts/unit-observer.py, a continuous observer for one measured unit that runs inside one cgroup, so a timed run can show that the host and the unit's own cgroup stayed within the admission limits throughout, not only at its boundaries.

What it checks, once a second plus one last sample at the stop: available memory (unrounded); host and workload memory and I/O full PSI per observation interval (unreadable PSI rejects); unbroken coverage on the monotonic clock; competing CPU as host busy cores minus the workload cgroup's own CPU over the whole run (--host-only counts the unit's own CPU); and, comparing every sample with the first, growth of memory.swap.current or memory.zswap.current, any change of zswpin, zswpout, zswpwb or of memory.events high, max, oom or oom_kill, a replaced directory, backward counters, unreadable required counters and a changed set of optional counters. Host swap-out and zswap activity is recorded per interval and rejects only with --storage-phase. It records its own cost against budgets and writes every sample to <output>.samples.jsonl. A clean summary does not qualify a run by itself; the protocol still declares observers, budgets, stop rules and retries.

Proof:

• Nine tests against a fake --proc-root and a fake cgroup directory: quiet clean, final sample at SIGTERM, OOM, swap, zswap and memory.stat changes, an excursion that rises and returns, missing or vanishing counters, pressure over the limit, a replaced directory, a backward counter, unreadable host I/O pressure (summary still written), competing CPU with and without a workload to subtract, --host-only, and --storage-phase. Weakening two checks (final-only events, unsubtracted host CPU) makes the corresponding tests fail.
• Real cgroups: a workload burning about 5 cores in its own scope came out clean with 3.5 competing cores; seven foreign loops outside the scope were rejected at 4.8 competing cores; observer cost was 0.003 cores and 19 MiB.
• bun run check passes.

Review: an independent review of the first version (GPT-6.1 Sol, xhigh) returned BLOCK with nine findings; all are fixed in the second commit. A second round could not run: the Codex quota is exhausted until Nov 2 and Muse's until Nov 2, so the rewrite has had no second local review. The private reviewer covers the head.

[bompus-ai-review]

Bompus AI Review could not run on 736ec70: no reviewer seat finished the review (GPT-6.1 Sol: usage limit reached; Muse Wonder Full: usage limit reached; SWE-2 Max: failed (exit 3)).

Nothing was reviewed. Pushing a new commit or re-running the review tries the seats again.

[bompus-ai-review]

Bompus AI Review (SWE-2 Max, max reasoning): 1 finding(s).

[bompus-ai-review]

Bompus AI Review (SWE-2 Max, max reasoning): 1 finding(s).

[bompus-ai-review]

Bompus AI Review could not run on c1dd437: no reviewer seat finished the review (GPT-6.1 Sol: usage limit reached; Muse Wonder Full: usage limit reached; SWE-2 Max: failed (exit 3)).

Nothing was reviewed. Pushing a new commit or re-running the review tries the seats again.

[bompus-ai-review]

Bompus AI Review (SWE-2 Max, max reasoning): 0 finding(s).

Review 3: only high findings and above are posted from here on.