django-logic 0.10.0

An overengineering sweep. ~1,900 lines removed across the engine, its tests and
its docs, four settings retired, and one addition: a system check for the
failure mode that motivated the audit. Every removal below was verified to have
no consumer — the two real consumers were searched under both import spellings
before anything was deleted.

Removed (breaking)

• RedisState and State.get_db_state(). RedisState was 199 of
  state.py's 326 lines and no consumer had ever set state_class to it;
  measured on a live deployment of 28 bound machines / 137 transitions, all 50
  process classes resolved to the base State. Use get_persisted_state() in
  place of get_db_state(). Two hardening issues turn out to have been spent on
  dead code — #139's token-in-key wrapper and #151's Lua compare-and-set — while
  the base-State ownership token and compare-and-delete unlock stay.
  Deliberately no RedisState = State alias: the two differ in locking
  semantics, so an alias would change behaviour silently where an ImportError
  says so plainly.

• LOG_KWARGS / LOG_KWARGS_REDACTOR. Scrub with a logging.Filter on the
  django-logic.transition logger instead. redact_log_kwargs keeps its
  shallow copy, which is a real fix, not a knob.

• PHASE2_STATE_GUARD = 'warn'. The phase-2 state guard now always
  enforces; a stale 'warn' is ignored. Its only purpose was restoring pre-0.4
  behaviour.

• SENTRY_TRANSACTION_NAMING. Transactions are always named.

• PROCESS_CLASS_ALIASES — removed rather than fixed, because it never did
  what it documented: the map was applied to the recorded bound class but
  _find_background_transition_in_owner compared owning_process_class
  verbatim, so renaming a nested owning process still yielded
  [unrestorable] with a correct alias configured. Drain in-flight rows
  before renaming a Process class — already the documented rule for the
  sibling action_name refactor.

• django_logic.conditions (all_related_in, any_related_in). The one
  consumer that wanted this pattern hand-wrote it four times with different
  empty-set semantics. docs/recipes/nested-processes.md now shows the closures
  inline.

• Process.get_transition_by_action_name and the public ignore_state=
  parameter of get_available_transitions. Neither had a caller in the library,
  its suite, or any consumer. To resolve one action, iterate
  get_available_transitions(user=..., action_name='x') and take the single
  match — that is what the removed method did. Gotcha: a stale call will not
  raise AttributeError pointing at this removal, because Process.__getattr__
  treats any non-underscore miss as an action name. Attribute access still
  returns a callable, and calling it reports an argument problem with a
  now-nonexistent action:

  TypeError: get_transition_by_action_name() accepts keyword arguments only
             (got 1 positional). Pass user and other values
  

  Grep for the name when upgrading rather than relying on the traceback.

• ProcessScenario.snapshot_on_failure and format_failure(snapshot=),
  never enabled by anyone in the 14 months since they shipped.
  snapshot() / from_snapshot() are unaffected.

• BaseTransition — folded into Transition, along with the
  conditions_class / failure_callbacks_class / failure_side_effects_class /
  next_transition_class swap hooks that had no override anywhere.
  Transition.is_background, side_effects_class, callbacks_class,
  permissions_class, conditions_class and Process.permissions_class all
  stay: consumers use them.

• ExecutionTracker is no longer re-exported from django_logic.testing
  (the class remains; track() is the supported path).

• process_name is now required on latest_message, message_for and
  uncompleted_message. The None default made it possible to silently
  reintroduce the #150 cross-process bleed by omitting an argument.

• JourneyStep.matches() — use ==.

• coverage_report no longer reads 0.8 or 0.9.0 log lines. Record a fresh
  log per run, which was always the documented practice.

• The non-finite-float pre-scan in serialize_kwargs. The
  json.dumps(allow_nan=False) round-trip rejects a superset (it also catches a
  non-finite float in a dict key, which the scan could not reach); the error
  message no longer names the offending key path.

• _infer_field_name: a TransitionMessage with no field_name now fails
  closed as [unrestorable] instead of guessing 'state', which could drive
  the wrong machine on a multi-process model.

Added

• django_logic.W002 — under BACKGROUND_EXECUTION='celery', reports any
  shipped safety-net task missing from the running app's beat schedule. This
  existed to catch a real incident: the README and beat_schedule()'s own
  docstring recommended app.conf.beat_schedule = {...}, which Celery silently
  ignores when the project also defines CELERY_BEAT_SCHEDULE in Django
  settings, and a consumer consequently ran seven weeks with all five tasks
  registered and never fired. Both now recommend
  app.conf['CELERY_BEAT_SCHEDULE'].
• django_logic.W003 — reports any DJANGO_LOGIC key this release removed.
  DJANGO_LOGIC has no unknown-key rejection, so every removal above fails
  open and silently; the sharpest case is a deployment that set
  LOG_KWARGS_REDACTOR for PII compliance, upgrades, and starts writing raw
  kwargs to its logs with no signal anywhere. The warning names the replacement
  for each key.

Fixed

• TaskCrashRedeliveryConfigTests hardcoded five task objects and so never
  covered recover_stranded_states; both acks_late and
  reject_on_worker_lost could be stripped from it with the suite green. The
  list is now derived from the module.
• The duplicated failure-side-effect savepoint in background/runner.py: #138
  had already moved that rule into commands.FailureSideEffects.execute with a
  stronger savepoint, and both call sites already ran inside the atomic block.

Docs

• Deleted five stale pre-implementation documents and the index that tiered
  them (2,311 lines), whose rot was already realized — docs/PLAN.md omitted
  five shipped settings while CLAUDE.md told engine-changers to read it first.
  The two still-open items from the Heroku-validation notes moved to TODO.md.
• De-duplicated the prose that README.md and CHANGELOG.md already owned, and
  collapsed the README testing chapter's overlap with docs/TESTING_GUIDE.md.
• CLAUDE.md now states the comment/docstring rule that reviewers had been
  citing as though it existed.

Repo

• Removed the nightly "stress" CI step (its STABILITY_STRESS_MODE was read by
  nothing, so it re-ran the identical tests), an unreferenced and broken
  docker-compose.test.yml service, and .cursor/rules/ (no audience, five
  false statements). The dev extra dropped pytest/pytest-django — there is
  no pytest configuration, so collection could never have worked — and
  djangorestframework, which no test imports. The drf extra stays.
• Test settings: one tests.dl_settings(**overrides) helper replaces 26
  hand-copied dicts, and 13 no-op @override_settings are gone.