Write and count bundle rows per package instead of holding them all

.github/workflows/ci.yml

@@ -81,9 +81,11 @@ jobs:
       - name: Test UK calibration source additions
         run: uv run pytest -q --durations=30 tests/test_chronicle_uk_national_2025.py
 
-      # GNU time reports the run's peak memory ("Maximum resident set size"): the
-      # merged bundle peaked near 15.6 GB in a local build, against the runner's
-      # 16 GB.
+      # GNU time reports the run's peak memory ("Maximum resident set size", in
+      # KiB). build_bundle writes and counts each package's rows as it reads them
+      # (chronicle#414), so this step peaked at 3.4 GiB for 1.74M facts on Nov 5,
+      # against 15.0 GiB, close to the runner's memory, when every row was held
+      # to the end.
       - name: Test merged Chronicle bundles
         run: |
           timer=""

changelog.d/bundle-streams-consumer-rows.fixed.md

@@ -0,0 +1,7 @@
+`build-bundle` writes and counts each source package's consumer rows as it reads
+them, and frees them before it builds the next package. It used to hold every
+package's rows until the end. CI's merged default bundle (1.74M rows) peaked at
+15.0 GiB that way and now peaks at 3.4 GiB; the UK suite's local build went from
+12.8 to 4.0 GiB. `consumer_facts.jsonl`, `coverage.json`, `source_packages.json`
+and the build report are unchanged (byte-identical on the UK suite); the facts
+file takes its name only once every package is in it.