DR original problem-management runtime lab, 2026-10-04

Run from project root:
python3 content/labs/problem-runtime/run.py --output /tmp/dr-problem-pool.json
Choose a new output filename to retain previous reports. Standard library only;
no networking, external accounts or platform database access. A private temporary
SQLite file contains one original synthetic settings row. Recorded CPython 3.13.1,
SQLite 3.51.2, Darwin arm64. Rolling Python 3.13 docs currently label 3.13.16;
source patch labels do not replace the runtime version reported in evidence.

Fifteen groups passed twice. Require exit 0, passed=15, failed=0, failure=null,
matching scriptSha256 and expected checks. Baseline reads return 17 and all three
leases. Faulty post-query exceptions retain one connection each, exhausting three;
the next attempt times out during acquisition without another SQL call. Recreating
the pool gives temporary capacity, but the same trigger exhausts it again. Capacity
five retains two free slots after three faults and none after five.

Corrected cleanup returns resources on exception and cancellation while preserving
their distinct outcomes. An await after acquisition but before try/finally leaves
a demonstrated cancellation gap. Cancellation of a still-waiting consumer does not
return someone else's lease. Eight tasks share three leases and all finish after
the gate opens, with maximum borrowed=3. Thirty success/exception/cancel cycles
return every lease after each cycle. A SQLite connection remains usable after its
transaction context exits, but not after explicit close. Replacing a closed leased
object restores the functional read; a positive free-slot count alone is not enough.
All owned tasks terminate, connections close and the private temporary tree is removed.

This is actual asyncio task/cancellation/Queue behavior and actual SQLite connection
use. Cancellation is delivered at the scripted await, not by interrupting a SQLite
query. The driver is synchronous in one event loop; this is not physical query
parallelism, an asynchronous database driver or a load benchmark. Short wait_for
limits terminate negative tests, not performance targets or SLAs. The original toy
pool omits production capabilities and must not be used as an enterprise library.
No WebSphere, JVM, JDBC, Oracle, remote database, production incident, ServiceNow/Jira
workflow, transaction recovery or human investigation workshop is executed.

See worksheet.txt for prediction, diagnosis and supplier-review decisions.
Representative product/workload/dependency tests, human practice, independent
specialist review and bilingual learner feedback remain outstanding.
