Codex-authored harness for the Cliff 2b investigation that surfaced the
"fragmentation-dominated, not pool growth" finding (PN12 flat at 137 MiB
across the cliff fire). Three files:
- instrument.py — observational monkey-patches at request/engine/worker
boundaries; writes CSV snapshots of allocator state + KV pool deltas
- sitecustomize.py — auto-loaded by Python at container start via
PYTHONPATH mount; gates instrumentation on RESIDENCY_LOG_PATH env var
- run-instrumented-soak.sh — wrapper that boots a compose with the
sidecar mounted, runs soak, joins instrumentation rows to soak turns
README.md (new) frames it as research-grade tooling, not a shipped feature:
explicit "no SLA" note, cross-link to #41 + docs/CLIFFS.md, soft warning
against running on production composes (sitecustomize slows boot).
Notes that results/residency-* is gitignored by repo policy — we don't
auto-publish diagnostic local results because they're large, rig-specific,
and the analytical value lives in issue threads + commit log + memory
entries that distill what was learned. Cross-rig contributors should
share findings via issue comments (summary tables), not raw CSV dumps.
This closes the last open thread from #217 (results/ cleanup decision).
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
