fix(b3a-deploy): memory request/limit + PYTHONUNBUFFERED on ingest Job

Bounded-window ingest peaks small; request 768Mi/limit 1536Mi makes the scheduler
place it with headroom and stops it evicting neighbours (a prior unbounded run hit
~2.3Gi on a shared 6.4Gi node). PYTHONUNBUFFERED streams progress logs live.

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

infra/k8s/jobs/fxhnt-backtest-ingest-job.yaml

@@ -37,8 +37,16 @@ spec:
           env:
             - name: FXHNT_BACKTEST_WAREHOUSE_PATH
               value: /backtest-data/warehouse.duckdb
+            - name: PYTHONUNBUFFERED         # stream ingest progress logs live (no stdout buffering)
+              value: "1"
             - name: TIINGO_API_KEY
               valueFrom: { secretKeyRef: { name: tiingo-api, key: key } }
+          # Bounded memory: the ingest now caps its in-flight window (~2x workers), so peak RSS is small and
+          # flat. The request+limit make the scheduler place this on a node with headroom and prevent it from
+          # evicting neighbours (an earlier unbounded run hit ~2.3Gi and was evicted off a shared 6.4Gi node).
+          resources:
+            requests: { memory: "768Mi", cpu: "250m" }
+            limits: { memory: "1536Mi" }
           volumeMounts:
             - { name: backtest-data, mountPath: /backtest-data }
       volumes: