DR original Change Management coordination lab, 2026-10-04

Run from project root:
python3 content/labs/change-coordination/run.py --output /tmp/dr-change-coordination.json
Choose a new output filename. Python standard library, Unix fcntl required.
Recorded on CPython 3.13.1, Darwin arm64, local filesystem. Rolling Python 3.13
documentation labels 3.13.16; archived Apple BSD manuals are historical references,
not a claim of a current Darwin specification. Actual observations are separate.

Thirteen groups passed twice. Require exit0, passed13, failed0, failure null,
matching scriptSha256 and workerSha256, and identical checks. The lab creates a
private temporary directory and invented JSON state. It starts five owned lock
workers plus a separate bypass writer, using controlled stdin/stdout interleaving.
The first owner acquires an exclusive nonblocking flock. A competitor exits75,
an application-defined code. A writes revision8. After release, B can acquire but
its revision7 plan is rejected. An explicitly reevaluated plan writes revision9;
repeating expected8 is rejected. Access acquisition does not reassess a plan.

With B alive, a nonparticipant writes revision10. B's later read detects this prior
change and rejects expected9. A deliberately blind write overwrites the intervening
value at revision11. Reading the revision and later using os.replace is NOT an
atomic conditional update against outside writers. No read/write-gap race is
claimed as executed: the bypass was deliberately completed before the read.
Cooperation on one stable local object is essential. No permissions are enforced.

The script sends SIGTERM only to an identified owned child and waits for exit.
A successor acquires the lock while revision11 remains; termination is not rollback.
It then deliberately unlinks its private lock pathname while an owner remains
active. Recreating the path gives another worker a different inode and a second
lock. This demonstrates an unsafe recovery pattern, not a production recommendation.
Never apply this unlink experiment to existing services. File age or existence does
not prove ownership or abandonment. All owned processes reach terminal status and
the temporary directory is removed, including on failure.

No distributed coordination, fencing, remote filesystem, database transaction,
remote task, real authorization, service manager, production incident, banking data,
external business effect, crash durability, power failure or benchmark. No human
workshop has been performed. This is not a production-ready coordination library.
Target acceptance must cover actual hosts, storage, every writer, intervention
authority and effects already initiated. Original fictional scenarios are not
procedures of financial institutions. Specialist and bilingual independent review remain pending.
