Prune each benchmark's weeks on its own runs

prune() took the newest run of each ISO week across the whole suite, so
a benchmark that ran that week, but not in its last run, lost the week:
with A on June 1 and 8 and B on June 7 and 14, A kept neither. Each
benched parametrization now keeps its own newest run of each week, and
its own MAX_RUNS; the 90-day window is still measured from the suite's
newest run, and every other row still keeps the suite's last two.

On the ci branch today that is 64 MB down to 8.3 MB, measured on the
files directly; the 13 MB 1d2f0a3 gave counted the latest run twice.

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

Author: hunhoffe

ci/scripts/merge_all.py

@@ -18,11 +18,13 @@
 def prune(rows, date_fmt="%Y-%m-%d %H:%M:%S"):
     """Keep the runs the reports and the Benchmarks page read.
 
-    A run is the rows of one date. A benched parametrization keeps every run of
-    the last `KEEP_DAYS` before the newest, the newest run of each ISO week
-    before that, and at most `MAX_RUNS` runs. The trends compare a run with
-    the one before it, so every other row keeps the last two runs. Measured
-    from the newest run, a skewed clock on one runner cannot empty the file.
+    A run is the rows of one date. Each benched parametrization keeps its own
+    runs of the last `KEEP_DAYS` before the suite's newest, its newest run of
+    each ISO week before that, and at most `MAX_RUNS` runs: a week another
+    test ran later in keeps this one's run too. The trends compare a run with
+    the one before it, so every other row keeps the suite's last two runs.
+    Measured from the newest run, a skewed clock on one runner cannot empty
+    the file.
     """
     dated = []
     for row in rows:
@@ -33,22 +35,27 @@ def prune(rows, date_fmt="%Y-%m-%d %H:%M:%S"):
     runs = sorted({date for date, _ in dated})
     if not runs:
         return []
-
     cutoff = runs[-1] - timedelta(days=KEEP_DAYS)
-    weekly = {}
-    for date in runs:
-        if date < cutoff:
-            weekly[date.isocalendar()[:2]] = date
-    kept = set(
-        sorted({d for d in runs if d >= cutoff} | set(weekly.values()))[-MAX_RUNS:]
-    )
     last_two = set(runs[-2:])
 
     benched = {row_key(row) for row in select_bench_rows(rows)}
+    runs_of = {}
+    for date, row in dated:
+        if row_key(row) in benched:
+            runs_of.setdefault(row_key(row), set()).add(date)
+    kept = {}
+    for key, dates in runs_of.items():
+        weekly = {}
+        for date in sorted(dates):
+            if date < cutoff:
+                weekly[date.isocalendar()[:2]] = date
+        recent = {date for date in dates if date >= cutoff}
+        kept[key] = set(sorted(recent | set(weekly.values()))[-MAX_RUNS:])
+
     return [
         row
         for date, row in dated
-        if date in last_two or (date in kept and row_key(row) in benched)
+        if date in last_two or date in kept.get(row_key(row), ())
     ]
 
 

iron/tests/infrastructure/test_merge_results.py

@@ -4,8 +4,8 @@
 """The results branch's retention, run as publish-results.yml runs it.
 
 ``merge_results.sh`` keeps a benched test's runs of the last 90 days, the
-newest of each ISO week before that, at most 400; and every other test's last
-two runs, all the trends read of it.
+newest of each ISO week it ran in before that, at most 400; and every other
+test's last two runs, all the trends read of it.
 """
 
 import csv
@@ -18,9 +18,10 @@
 COLUMNS = ["Commit", "Date", "Test Path", "Test", "Checks", "Bench", "Latency (mean)"]
 BENCHED = ("iron/tests/operators/test_catalog.py::test_operator", "npu2-GEMV-M_2048")
 PLAIN = ("iron/tests/common/test_graph.py::test_record", "test_record")
+OTHER = ("iron/tests/operators/test_catalog.py::test_operator", "npu2-GEMV-M_4096")
 
 
-def _rows(dates):
+def _rows(dates, tests=((BENCHED, "yes"), (PLAIN, "no"))):
     return [
         {
             "Commit": "abc1234",
@@ -32,7 +33,7 @@ def _rows(dates):
             "Latency (mean)": "1.0",
         }
         for date in dates
-        for (path, test), bench in ((BENCHED, "yes"), (PLAIN, "no"))
+        for (path, test), bench in tests
     ]
 
 
@@ -74,3 +75,14 @@ def test_a_benched_test_keeps_at_most_400_runs(tmp_path):
     hours = [newest - timedelta(hours=n) for n in range(1, 500)]
     kept = _merge(tmp_path, _rows(hours), _rows([newest]))
     assert kept[BENCHED[1]] == {newest} | set(hours[:399])
+
+
+def test_each_benchmark_keeps_its_own_newest_run_of_a_week(tmp_path):
+    """A week another benchmark ran later in keeps this one's run too."""
+    a, b = ((BENCHED, "yes"),), ((OTHER, "yes"),)
+    june = [datetime(2026, 6, d, 10) for d in (1, 7, 8, 14)]
+    newest = datetime(2026, 10, 30, 10)
+    history = _rows([june[0], june[2]], a) + _rows([june[1], june[3]], b)
+    kept = _merge(tmp_path, history, _rows([newest], a + b))
+    assert kept[BENCHED[1]] == {june[0], june[2], newest}
+    assert kept[OTHER[1]] == {june[1], june[3], newest}