fix(ingest): address review round 1 (reclaim queue + tests)

- Register the periodic stalled-job reclaim on a dedicated ingest-maintenance
  queue that every worker drains (any --tier), so reclaim still fires when the
  fast fleet is scaled to zero and only ocr workers run. procrastinate's
  periodic-defer dedup keeps it single-run across drainers.
- escalation: mark `unsupported`/`forced` reason labels as reserved (not raised).
- processor: note that options/progress_callback are intentionally not threaded
  through _parse_pdf_tier yet (symmetric with the inline path).
- tests: assert TieredEscalationStrategy backoff progression (4/8/16/…/300s);
  cover get_ingest_pending per-queue aggregation + the legacy job_counts
  fallback; add an external-path zero-page no-escalation case; use the canonical
  INGEST_QUEUE_FAST instead of the back-compat alias.

Deck #323.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

tests/unit/vector/test_ingest_status.py

@@ -11,21 +11,45 @@ pytestmark = pytest.mark.unit
 
 
 class TestGetIngestPending:
-    async def test_postgres_reads_job_counts(self):
+    async def test_postgres_aggregates_by_queue(self):
+        """Per-queue counts are summed into fleet-wide totals (Deck #323)."""
         producer = AsyncMock()
-        producer.job_counts.return_value = {"todo": 5, "doing": 2, "failed": 1}
+        producer.job_counts_by_queue.return_value = {
+            "ingest-fast": {"todo": 5, "doing": 1},
+            "ingest-ocr": {"doing": 1, "failed": 1},
+        }
 
         result = await get_ingest_pending(
             task_producer=producer,
             document_receive_stream=None,
             ingest_queue="postgres",
         )
-        assert result.pending == 7  # todo + doing
+        assert result.pending == 7  # todo(5) + doing(1+1)
         assert result.job_counts == {"todo": 5, "doing": 2, "failed": 1}
+        assert result.job_counts_by_queue == {
+            "ingest-fast": {"todo": 5, "doing": 1},
+            "ingest-ocr": {"doing": 1, "failed": 1},
+        }
+
+    async def test_postgres_falls_back_to_aggregate_counts(self):
+        """A producer without job_counts_by_queue uses the aggregated call."""
+
+        class LegacyProducer:
+            async def job_counts(self):
+                return {"todo": 5, "doing": 2, "failed": 1}
+
+        result = await get_ingest_pending(
+            task_producer=LegacyProducer(),
+            document_receive_stream=None,
+            ingest_queue="postgres",
+        )
+        assert result.pending == 7
+        assert result.job_counts == {"todo": 5, "doing": 2, "failed": 1}
+        assert result.job_counts_by_queue is None
 
     async def test_postgres_degrades_to_zero_on_error(self):
         producer = AsyncMock()
-        producer.job_counts.side_effect = RuntimeError("db down")
+        producer.job_counts_by_queue.side_effect = RuntimeError("db down")
 
         result = await get_ingest_pending(
             task_producer=producer,

tests/unit/vector/test_procrastinate_producer.py

@@ -48,7 +48,8 @@ class TestProcrastinateTaskProducer:
         assert len(jobs) == 1
         job = jobs[0]
         assert job["task_name"] == pq.INGEST_TASK_NAME
-        assert job["queue_name"] == pq.INGEST_QUEUE_NAME
+        # New jobs are deferred onto the cheapest tier's queue.
+        assert job["queue_name"] == pq.INGEST_QUEUE_FAST
         assert job["queueing_lock"] == "alice:note:42"
         assert job["lock"] is None  # no execution lock (crash-deadlock guard)
         assert job["args"]["doc_id"] == "42"

tests/unit/vector/test_tiered_escalation_strategy.py

@@ -5,6 +5,8 @@ and the procrastinate TieredEscalationStrategy that turns a raised exception
 into a queue-hop / same-tier retry / give-up decision.
 """
 
+from datetime import datetime, timezone
+
 import httpx
 import pytest
 from procrastinate.jobs import Job
@@ -87,6 +89,20 @@ class TestTieredEscalationStrategy:
         assert decision.queue is None
         assert decision.retry_at is not None
 
+    def test_transient_backoff_progression(self):
+        # min(4 * 2**(attempts-1), 300): 4, 8, 16, ... capped at 300s.
+        strat = self._strategy(max_transient=100)
+        for attempts, expected in [(1, 4), (2, 8), (3, 16), (4, 32), (20, 300)]:
+            decision = strat.get_retry_decision(
+                exception=httpx.ConnectError("x"), job=_job(attempts=attempts)
+            )
+            assert decision is not None and decision.retry_at is not None
+            delta = (decision.retry_at - datetime.now(timezone.utc)).total_seconds()
+            # retry_at = now + wait; allow a small window for execution time.
+            assert expected - 2 <= delta <= expected + 1, (
+                f"attempts={attempts}: delta={delta:.2f}s, expected≈{expected}s"
+            )
+
     def test_transient_gives_up_over_cap(self):
         decision = self._strategy(max_transient=5).get_retry_decision(
             exception=httpx.ConnectError("refused"), job=_job(attempts=5)