[PhiLily]

benchmarks/killbench.py gains --external-postgres. Without it,
orchestrate() calls bench.ensure_docker() and bench.ensure_container(),
which manage a local Docker Desktop container through docker --context desktop-linux, and a GitHub Actions runner has neither that context nor
a docker binary. The flag skips exactly those two calls. The rest of the
setup, ensure_databases(), migrate(), truncate() and
collect_environment(), goes through plain psycopg or Django and works
against any PostgreSQL reachable at bench.PG (127.0.0.1:54330, user
postgres, password ox), the values benchsite/settings_ox.py and its
kill-test derivatives already hard-code, so a service container that
publishes that port needs no settings change. Without the flag the
harness runs exactly as before.

.github/workflows/kill-smoke.yml is new. Its smoke job starts a
postgres:16 service container on port 54330, installs the checked-out
package with Django~=6.1.0 and django-tasks-db==0.13.0, and runs
killbench.py --smoke --external-postgres. django-tasks-db is in no
dependency group: the harness's tasksdb arm imports it directly, and the
explicit pin keeps an upstream db_worker flag change from failing this
job without a deliberate bump here. A second step reads the
results-kill-*-SMOKE.json the harness writes and fails the job if either
arm's harness_ok_all is false, or if in the django-ox arm any task never
reached a terminal state or ended FAILED or LOST. Those three counts are
not checked for the tasks-db arm: it leaves RUNNING rows behind by
design, as the README's "Worker death" disclosures record, so a check
over both arms could never pass. Results and worker logs upload as an
artifact whether the job passes or fails, matched by glob, so the
harness's own -SMOKE filename suffix is kept.

The job runs on a weekly schedule, Wednesday 05:00 UTC ("0 5 * * 3"),
clear of this repository's other two schedules at 03:17 daily and 06:00
on Mondays; on workflow_dispatch; and on a pull request that touches
what the job depends on: the workflow file itself, killbench.py,
bench.py and benchsite/**. The job reads no github.event.pull_request.*
or other pull-request-only expression, so a green run means the same
thing under any of the three triggers.

benchmarks/README.md: the opening line of the "Worker death" section
said the harness reuses bench.py for the container; it now says so for
the default path only and names --external-postgres for the other one.
A new bullet in that section's disclosure list names the weekly workflow
and states that a failing scheduled run reaches only whoever last
committed to the workflow file; nothing else in this repository watches
it. A new "Reproduce this yourself" subsection gives the commands for
the smoke run and the full run, where each writes its output, and how to
submit a run from another host: attach the JSON to an issue or pull
request here.

Nothing in django-ox itself changes, only how its worker-death harness
is exercised and reported on. benchmarks/ is not shipped in the package,
sits outside the ruff and mypy configuration, and gets no CHANGELOG
entry.

Closes #145