…ing + typing/dedup/docs/tests

Decision (with reviewer): reaper-as-safety-net for the un-catchable strand
windows + fix what's catchable + document + gate on PR3.

Failure handling:
- [#69 Critical] run_batch_with_barrier wraps BOTH work + decrement in the abort:
  a decrement-side failure (guard / DB / last-batch callback dispatch) tears the
  barrier down in-body instead of stranding to expiry.
- [#79] extracted _abort_barrier_in_body — logs when the teardown itself fails
  (was silently suppressed under a misleading "torn down" message).
- [#74/#81] documented the two un-catchable strand windows (hard-crash-during-work,
  post-commit callback-dispatch-fail) as a HARD reaper dependency for PR3.
- [#86] finalise cleanup split into independent try/excepts with distinct logs.

Typing / clarity:
- [#1] BarrierContext(TypedDict) for _barrier_context (header fan-out, run_batch_with_barrier,
  process_file_batch).
- [#3] renamed CallbackDescriptor "backend" -> "transport" (WorkflowTransport value;
  avoids the QueueBackend "pg" collision).
- [#27] is_pg_transport() predicate in core; used in orchestration_utils + pg_barrier.
- [#20] extracted _dispatch_pg() — single home for cycle-avoiding local import + backend=PG.
- [#35] normalize_transport() at the general worker entry (parity w/ api/scheduler).
- [#94] log when a header has no queue option.
- [#9/#13] fixed born-stale comment + kwargs-not-args docstring.

Tests (+#37/#41): last-batch self-chains callback to PG + cleans up barrier/dedup;
decrement-failure aborts; PG-branch mid-loop dispatch-failure deletes row; header
args/queue/pre-existing-kwargs preservation. 137 barrier/dedup/routing tests green;
bootstrap clean under WORKER_BARRIER_BACKEND=pg.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…fire-and-forget) (#2069)

* UN-3563 [FEAT] PG Queue 9e PR 2c — live PG fan-out/barrier/callback (fire-and-forget)

Wires the coupled pipeline's fan-out → barrier → callback onto the PG queue for
a transport=="pg_queue" execution. Gated: resolve_transport() still returns
celery (PR3 Flipt flips it), so the whole PG branch is present-but-unreachable —
default path byte-identical. Orchestrator task (async_execute_bin) stays on
Celery (hybrid); routing it onto PG is a 2d follow-up.

- barrier.py: Barrier Protocol + CeleryChordBarrier/RedisDecrBarrier accept (and
  ignore) a `transport` param; CallbackDescriptor gains an optional `backend`.
- orchestration_utils._barrier_for_transport: pg_queue → fresh PgBarrier()
  (bypasses the WORKER_BARRIER_BACKEND singleton), else the singleton.
- pg_barrier.PgBarrier.enqueue(transport): pg_queue → fire-and-forget mode —
  _dispatch_header_pg sends each header via dispatch(backend=PG) with an injected
  _barrier_context {execution_id, batch_index, callback_descriptor}, no .link;
  descriptor marked backend=pg_queue; UPSERT block also clears pg_batch_dedup
  (greptile #2068 reuse-reset). _fire_barrier_callback self-chains the callback
  onto PG when backend==pg_queue. clear_execution_batches at finalise + abort.
  run_batch_with_barrier(): claim → work → in-body _barrier_pg_decrement;
  redelivery skips; exception → barrier_pg_abort.
- file_processing.process_file_batch(_barrier_context=None): core routes None →
  _run_batch_stages (celery chord path), else → run_batch_with_barrier.
- general/api fan-outs thread transport into create_chord_execution.

Tests: +8 PgBarrier fire-and-forget + 2 orchestration routing + 2 process_file_batch
routing. Each test file green alone; ruff clean. End-to-end forced-pg dev-test
pending before PR.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* UN-3563 fix SonarCloud S1172: drop unused task_instance from _run_batch_stages

The extracted _run_batch_stages never uses task_instance — its only purpose
(deriving celery_task_id) happens in _process_file_batch_core before the call.
Removed the param + updated both call sites. _process_file_batch_core keeps
task_instance (it reads .request.id). Routing test mocks with *a, unaffected.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* UN-3563 address review (muhammad-ali-e, 15): strand-on-failure hardening + typing/dedup/docs/tests

Decision (with reviewer): reaper-as-safety-net for the un-catchable strand
windows + fix what's catchable + document + gate on PR3.

Failure handling:
- [#69 Critical] run_batch_with_barrier wraps BOTH work + decrement in the abort:
  a decrement-side failure (guard / DB / last-batch callback dispatch) tears the
  barrier down in-body instead of stranding to expiry.
- [#79] extracted _abort_barrier_in_body — logs when the teardown itself fails
  (was silently suppressed under a misleading "torn down" message).
- [#74/#81] documented the two un-catchable strand windows (hard-crash-during-work,
  post-commit callback-dispatch-fail) as a HARD reaper dependency for PR3.
- [#86] finalise cleanup split into independent try/excepts with distinct logs.

Typing / clarity:
- [#1] BarrierContext(TypedDict) for _barrier_context (header fan-out, run_batch_with_barrier,
  process_file_batch).
- [#3] renamed CallbackDescriptor "backend" -> "transport" (WorkflowTransport value;
  avoids the QueueBackend "pg" collision).
- [#27] is_pg_transport() predicate in core; used in orchestration_utils + pg_barrier.
- [#20] extracted _dispatch_pg() — single home for cycle-avoiding local import + backend=PG.
- [#35] normalize_transport() at the general worker entry (parity w/ api/scheduler).
- [#94] log when a header has no queue option.
- [#9/#13] fixed born-stale comment + kwargs-not-args docstring.

Tests (+#37/#41): last-batch self-chains callback to PG + cleans up barrier/dedup;
decrement-failure aborts; PG-branch mid-loop dispatch-failure deletes row; header
args/queue/pre-existing-kwargs preservation. 137 barrier/dedup/routing tests green;
bootstrap clean under WORKER_BARRIER_BACKEND=pg.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* UN-3563 fix run_batch_with_barrier strand-window doc inconsistency (review)

The second "NOT catchable" bullet conflated two different things: it described
the in-body catchable abort ("the abort here removes the row") and a *software*
callback-dispatch failure — but that failure is already caught + torn down by
step 3's wrap (paragraph 1), so it doesn't belong under the un-catchable
heading, and on the PG path _fire_barrier_callback IS the enqueue so "committed
but before the enqueue" couldn't both hold.

Rewrote the bullet to the genuinely un-catchable window: a hard crash BETWEEN
the decrement committing (remaining→0) and the callback enqueue completing —
decrement committed (redelivery blocked by the marker), process gone before the
callback enqueues or any abort runs, row survives to expiry, reaper-only
recovery. Explicitly notes a software dispatch failure is the catchable case.
Keeps this list an accurate spec for the PR-3 reaper-recovery dependency.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* UN-3563 address greptile (#2069, 2): clear dedup on mid-loop PG failure + carry fairness on PG callback

Both in the gated PG path (greptile 4/5, safe to merge).

- Issue 1: PgBarrier.enqueue mid-loop dispatch-failure handler now also calls
  clear_execution_batches on the PG path. Earlier headers may have committed a
  claim_batch marker; with the barrier row deleted, their in-flight
  barrier_pg_abort is a no-op (already_aborted) and never reaches the clear
  inside it, so reclaim the markers directly here.
- Issue 2: the PG callback now carries the producer's fairness. Added
  _fairness_from_headers() to reconstruct the FairnessKey from the stored
  x-fairness-key headers and pass it to _dispatch_pg, so the callback rides the
  same org/priority as the Celery path (was always default priority).

Tests: +fairness-carried / +fairness-none-safe on _fire_barrier_callback;
extended the PG mid-loop test to assert an already-claimed marker is reclaimed.
75 barrier/dedup tests green; bootstrap clean under WORKER_BARRIER_BACKEND=pg.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* UN-3563 fix SonarCloud S3776: reduce PgBarrier.enqueue cognitive complexity (17→under 15)

Extracted the per-header dispatch loop into PgBarrier._dispatch_headers — the
deeply-nested for→try/except→if/else→if (PG-vs-celery branch + mid-loop failure
teardown + PG dedup-clear) was the complexity driver. enqueue now calls the
helper; behaviour identical. radon: enqueue C(11)→B(6); ruff C901 passes.
75 barrier/dedup tests green; ruff + ruff-format clean.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* UN-3563 fix greptile #2069: mid-loop dedup-clear test passed for the wrong reason

The pre-seeded claim_batch marker was wiped by enqueue's UPSERT block (the
reuse-reset DELETE) before the dispatch loop, so the mid-loop
clear_execution_batches deleted 0 rows — the count==0 assertion passed on the
UPSERT, not the guard under test. Now the first dispatch side-effect claims the
marker AFTER the UPSERT (simulating a fast PG consumer), so the mid-loop clear
is what removes it. Verified: with the clear disabled the marker orphans (count=1).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>