Update the FOMC events dataset

PIPELINE.md

@@ -221,15 +221,29 @@ change, so a run that already holds the row does not ask TreasuryDirect for it
 again: the first build reads 6 229 artifacts in about thirty-five minutes, and
 every run after it reuses them and finishes in **four seconds**.
 
-Only rows whose instant came from an artifact are reused. A row dated to the
-end of its day is one whose artifact was never found, and the next run looks
-again in case the Treasury has since published one. The rate feeds are reused
-by whole year, and never the current one. `--rebuild` ignores held rows
-entirely, which is what a parser change needs.
-
-Between them, the two reuse paths take a run that fetched 12 832 documents down
-to **seven requests** and about **three seconds**, producing byte-identical
-tables.
+What decides reuse for an auction is `raw_sha256`, not the precision of its
+instant. A row carrying a digest is one whose artifact was fetched and read,
+and that is settled whether or not the artifact turned out to state a release
+time -- **1 357 of them do not**, and reusing only the `minute` rows meant
+refetching every one of those on every run, forever, looking for a line that
+was never going to appear. A row with no digest is one whose artifact was never
+found, and that one is worth looking for again.
+
+For the rate feeds, **every** past year is settled, not only the ones that
+produced rows. The real curve begins in 2003 and the years before it return a
+file with nothing but a header; holding only the years with data meant fetching
+thirteen empty years on every run to be told again that they are empty.
+
+Five auctions of 11 114 state a noncompetitive closing time in the artifact
+that the query leaves empty, so those two fields are carried forward from the
+held announcement rather than rebuilt from the query. Without that the
+incremental build would quietly drop a field the dataset already had.
+
+Between them the reuse paths take a run that fetched 12 832 documents down to
+**seven requests** and about **two and a half seconds**, and the tables it
+produces are identical to the full build's in all eight configs.
+
+`--rebuild` ignores held rows entirely, which is what a parser change needs.
 
 ## Scheduling
 

data/auction_announcements/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:692a186e39738bc86fc3b8af18636588db19d5913078fcf7bb0d688ae829f639
-size 484959
+oid sha256:a7e6cbb7ac9ef1e08b451fc205789b587a211cfacc6ae1fd688e7699f28f26d5
+size 479469

data/auction_calendar/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fb441c1e6dc9a94a129f12ffa42d49af122fcd8f100b59f237f5e44fa12cbf6d
-size 423326
+oid sha256:9f7e6d146262f9d5c07231213be3c1210956ae538d32d2f54f3ba7d75abfa810
+size 417766

data/auction_results/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f5b44ea9b34f5e17eb57ce07f91fc145e5a05fcf8946c6644e69140648ce9d68
-size 1622007
+oid sha256:26a7eaac97521473015ead61d5b1d0994eeb376c7d73e511970b3cdee49721a4
+size 1618329

data/bill_rates/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c064c75cb56860cd1ed0079288433a2bb4d04b974caa7063c5f49fff2caedb5e
-size 367317
+oid sha256:0a4fca2ba6f4c3c23769fb7f74f3897875f068ac951166bb215d5c5a3144d0b5
+size 368155

data/curve_features/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0d770c3a9de413c0254eae39e571baf343f2e496e9cf0531d8b397a4d4f307cb
+oid sha256:9d3f4c480ed4beb53570b75ac55b9fb353f9fcb63749c85d5051cd1552a0bad8
 size 189724

data/nominal_yield_curve/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9a936baed49df2ea2fc8f817ef40cc4c3298159885fd818ee51bb6029fea916f
-size 716156
+oid sha256:d3add81c0817d90cbe5d9e97d77cc7326fe79b8203395cbd4df31618ebfe28d7
+size 716704

data/pit/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cc259aa2cd92eaba61469855685332dec9b7dfe273784080a2f0431eaeac4a31
-size 2717527
+oid sha256:2948d11c2fc9a90792fbfa22d3d0a33f4ff4b98bf2aaa5a9dbcfe63ca5736eb1
+size 2724267

data/pit/us_treasury_events.delta/_delta_log/00000000000000000000.json

@@ -1,4 +1,4 @@
-{"commitInfo":{"timestamp":1789281137938,"operation":"WRITE","operationParameters":{"mode":"Overwrite"},"engineInfo":"delta-rs:py-1.6.3","operationMetrics":{"num_added_files":1,"num_removed_files":0,"num_partitions":0,"num_added_rows":268960,"execution_time_ms":154},"clientVersion":"delta-rs.py-1.6.3"}}
+{"commitInfo":{"timestamp":1789281131033,"operation":"WRITE","operationParameters":{"mode":"Overwrite"},"engineInfo":"delta-rs:py-1.6.2","operationMetrics":{"execution_time_ms":178,"num_added_files":1,"num_added_rows":268960,"num_partitions":0,"num_removed_files":0},"clientVersion":"delta-rs.py-1.6.2"}}
 {"protocol":{"minReaderVersion":1,"minWriterVersion":2}}
-{"metaData":{"id":"b4b79ff4-fa82-4a52-a5c7-969fe63f41bf","name":null,"description":null,"format":{"provider":"parquet","options":{}},"schemaString":"{\"type\":\"struct\",\"fields\":[{\"name\":\"entity_id\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"event_date\",\"type\":\"date\",\"nullable\":true,\"metadata\":{}},{\"name\":\"knowledge_date\",\"type\":\"timestamp\",\"nullable\":true,\"metadata\":{}},{\"name\":\"subset\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"record_id\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"actual\",\"type\":\"double\",\"nullable\":true,\"metadata\":{}},{\"name\":\"previous\",\"type\":\"double\",\"nullable\":true,\"metadata\":{}},{\"name\":\"change\",\"type\":\"double\",\"nullable\":true,\"metadata\":{}},{\"name\":\"forecast\",\"type\":\"double\",\"nullable\":true,\"metadata\":{}},{\"name\":\"unit\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"value_method\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"knowledge_time_precision\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"source_url\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"parser_version\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"ingested_at\",\"type\":\"timestamp\",\"nullable\":true,\"metadata\":{}}]}","partitionColumns":[],"createdTime":1789281137783,"configuration":{}}}
-{"add":{"path":"part-00000-2bb0a7ee-ab33-41bf-9c32-fe0953ae83dd-c000.snappy.parquet","partitionValues":{},"size":4421397,"modificationTime":1789281137938,"dataChange":true,"stats":"{\"numRecords\":268960,\"minValues\":{\"subset\":\"auction_results\",\"value_method\":\"derived\",\"event_date\":\"1979-10-31\",\"unit\":\"percent\",\"change\":-98.69,\"knowledge_date\":\"1979-11-01T04:59:59Z\",\"knowledge_time_precision\":\"date\",\"record_id\":\"UST_1979-10-31_912827KC5_RES\",\"actual\":-1.91,\"previous\":-1.91,\"entity_id\":\"UST_AUCTION_101D_CMB_ALLOTMENT_AT_HIGH_PCT\",\"source_url\":\"https://home.treasury.gov/resource-center/data-chart-center/inte\",\"parser_version\":\"1.0.0\",\"ingested_at\":\"2026-09-13T06:32:16.488770Z\"},\"maxValues\":{\"ingested_at\":\"2026-09-13T06:32:16.488770Z\",\"subset\":\"real_yield_curve\",\"value_method\":\"reported\",\"unit\":\"ratio\",\"entity_id\":\"UST_REAL_LT_REAL_AVERAGE\",\"actual\":100.0,\"record_id\":\"real_2026-09-11_7Y\",\"knowledge_time_precision\":\"modeled_conservative\",\"knowledge_date\":\"2026-09-11T22:00:00Z\",\"event_date\":\"2026-09-11\",\"previous\":100.0,\"parser_version\":\"1.0.0\",\"source_url\":\"https://www.treasurydirect.gov/xml/R_20260910_3.xml\",\"change\":98.15},\"nullCount\":{\"source_url\":0,\"entity_id\":0,\"record_id\":0,\"value_method\":0,\"event_date\":0,\"parser_version\":0,\"ingested_at\":0,\"forecast\":268960,\"subset\":0,\"actual\":0,\"knowledge_date\":0,\"knowledge_time_precision\":0,\"previous\":1287,\"unit\":0,\"change\":1287}}","tags":null,"baseRowId":null,"defaultRowCommitVersion":null,"clusteringProvider":null}}
+{"metaData":{"id":"f3316422-d5b5-4edf-a62e-452af79dbd56","name":null,"description":null,"format":{"provider":"parquet","options":{}},"schemaString":"{\"type\":\"struct\",\"fields\":[{\"name\":\"entity_id\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"event_date\",\"type\":\"date\",\"nullable\":true,\"metadata\":{}},{\"name\":\"knowledge_date\",\"type\":\"timestamp\",\"nullable\":true,\"metadata\":{}},{\"name\":\"subset\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"record_id\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"actual\",\"type\":\"double\",\"nullable\":true,\"metadata\":{}},{\"name\":\"previous\",\"type\":\"double\",\"nullable\":true,\"metadata\":{}},{\"name\":\"change\",\"type\":\"double\",\"nullable\":true,\"metadata\":{}},{\"name\":\"forecast\",\"type\":\"double\",\"nullable\":true,\"metadata\":{}},{\"name\":\"unit\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"value_method\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"knowledge_time_precision\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"source_url\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"parser_version\",\"type\":\"string\",\"nullable\":true,\"metadata\":{}},{\"name\":\"ingested_at\",\"type\":\"timestamp\",\"nullable\":true,\"metadata\":{}}]}","partitionColumns":[],"createdTime":1789281130859,"configuration":{}}}
+{"add":{"path":"part-00000-b2b0b029-9fa3-4076-932c-17ffc313e68b-c000.snappy.parquet","partitionValues":{},"size":4422313,"modificationTime":1789281131031,"dataChange":true,"stats":"{\"numRecords\":268960,\"minValues\":{\"unit\":\"percent\",\"subset\":\"auction_results\",\"event_date\":\"1979-10-31\",\"knowledge_date\":\"1979-11-01T04:59:59Z\",\"change\":-98.69,\"value_method\":\"derived\",\"record_id\":\"UST_1979-10-31_912827KC5_RES\",\"source_url\":\"https://home.treasury.gov/resource-center/data-chart-center/inte\",\"knowledge_time_precision\":\"date\",\"actual\":-1.91,\"parser_version\":\"1.0.0\",\"ingested_at\":\"2026-09-13T06:32:10.260494Z\",\"previous\":-1.91,\"entity_id\":\"UST_AUCTION_101D_CMB_ALLOTMENT_AT_HIGH_PCT\"},\"maxValues\":{\"unit\":\"ratio\",\"parser_version\":\"1.0.0\",\"entity_id\":\"UST_REAL_LT_REAL_AVERAGE\",\"subset\":\"real_yield_curve\",\"record_id\":\"real_2026-09-11_7Y\",\"previous\":100.0,\"knowledge_time_precision\":\"modeled_conservative\",\"knowledge_date\":\"2026-09-11T22:00:00Z\",\"value_method\":\"reported\",\"actual\":100.0,\"change\":98.15,\"ingested_at\":\"2026-09-13T06:32:10.260494Z\",\"source_url\":\"https://www.treasurydirect.gov/xml/R_20260910_3.xml\",\"event_date\":\"2026-09-11\"},\"nullCount\":{\"entity_id\":0,\"knowledge_date\":0,\"value_method\":0,\"unit\":0,\"ingested_at\":0,\"previous\":1287,\"forecast\":268960,\"parser_version\":0,\"event_date\":0,\"subset\":0,\"actual\":0,\"change\":1287,\"knowledge_time_precision\":0,\"record_id\":0,\"source_url\":0}}","tags":null,"baseRowId":null,"defaultRowCommitVersion":null,"clusteringProvider":null}}

data/pit/us_treasury_events.delta/{part-00000-2bb0a7ee-ab33-41bf-9c32-fe0953ae83dd-c000.snappy.parquet → part-00000-b2b0b029-9fa3-4076-932c-17ffc313e68b-c000.snappy.parquet}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2fd230b729a22d09c271aa25659ed5089d8a4a9c0b86adf9f670243cf3825f3c
-size 4421397
+oid sha256:0622c8b9a7af158cf88ebca5c9c5c961d30ef4c4010cd67fc53ac1f06143fdf9
+size 4422313

data/real_yield_curve/part-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7a8e2dba4f8faa9305d0889cb1a79a522350c5eb45669424223bc57cce415336
-size 404917
+oid sha256:574beeb24aa0a7952d8e92960e8ad9296aea153b1f0c37f1622fd932ad95e664
+size 405997

manifest.json

@@ -278,7 +278,7 @@
   "event_date": "event_date",
   "event_date_type": "date",
   "frequency_seconds": null,
-  "generated_at": "2026-09-13T06:32:18.135320+00:00",
+  "generated_at": "2026-09-13T06:32:11.129328+00:00",
   "knowledge_date": "knowledge_date",
   "knowledge_date_policy": {
     "auction_results": "the release minute stated in the Treasury's own result artifact, from March 2008; the end of the auction day before that",
@@ -325,6 +325,6 @@
     "build_version": "1.0.0",
     "package_version": "1.0.0",
     "parser_version": "1.0.0",
-    "recipe_hash": "sha256:60df5049dcd4040bdb200538634016989252ecf3b34869851220140908e55608"
+    "recipe_hash": "sha256:2807f605f7c9b8a56e579329be5a28fc38539806641e82030a079a47db686251"
   }
 }

metadata/build.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:65b64053deb2eaab79fa9b57d43f7654bd2b06c09143ce999a26f39ebf48ddd3
-size 4221
+oid sha256:687437599d62c0354dd6e0c60bf35873431cdfdaa9ef8a5e28a95e0a23eec6a2
+size 5400

metadata/coverage.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:205f248a45dbebb9c7080d6ce6aaecdaf31e36f551603e788d32ec4f912dfcd9
-size 4481
+oid sha256:1ec7f64315ba94eadf641deefff9edc8cc6d9976847a0968fca35a3fc9d24b92
+size 5293

metadata/field_coverage.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:801f6bc84e6261ed0df3fad147b5155560db2469904d4b611fef0445937d3634
+oid sha256:fcdd257538b3ab9b7b50e10926538138ba3d43e8389b0ecb86abf83b2754486f
 size 3288

metadata/yield_curve_methodology.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:882bffde7a4e2b4fde94cc44888ae9b4b828527dec37378fcf990c5cd99ede31
-size 3984
+oid sha256:6d6f5f4f319e27c751f43091c907c3055e12a58da5a737449ceb646a75fd7462
+size 7308

quality_report.json

@@ -1,6 +1,6 @@
 {
   "ok": true,
-  "checked_at": "2026-09-13T06:32:17.960289+00:00",
+  "checked_at": "2026-09-13T06:32:11.067792+00:00",
   "counts": {
     "auction_calendar": 11114,
     "auction_announcements": 11114,

recipe/pipeline.py

@@ -30,7 +30,7 @@ from pathlib import Path
 
 import polars as pl
 
-from .config import COUNTRY, Settings, validate_user_agent
+from .config import COUNTRY, FIRST_CURVE_YEAR, Settings, validate_user_agent
 from .coverage import write_build_record, write_coverage
 from .http import Artifact, Fetcher
 from .methodology import methodology_for, write_registry
@@ -332,7 +332,32 @@ def result_row(auction: Auction, *, artifact: Artifact | None, times: dict) -> d
 CURVE_CONCURRENCY = 6
 
 
-def held_curve_years(data_dir: Path) -> dict[str, set[int]]:
+def held_closing_times(data_dir: Path) -> dict[str, dict]:
+    """The closing times a previous run read out of an artifact.
+
+    Almost every auction has its closing times in the query as well, but five
+    of 11 114 do not: the artifact states a noncompetitive close that the query
+    leaves empty. Rebuilding those announcements from the query alone would
+    quietly drop a field that the dataset already had, and the incremental
+    build would stop matching the full one.
+    """
+
+    try:
+        rows = read_table(data_dir, "auction_announcements").collect()
+    except (FileNotFoundError, pl.exceptions.ComputeError):
+        return {}
+    if rows.is_empty():
+        return {}
+    return {
+        row["auction_id"]: {
+            "competitive_close_at": row["competitive_close_at"],
+            "noncompetitive_close_at": row["noncompetitive_close_at"],
+        }
+        for row in rows.iter_rows(named=True)
+    }
+
+
+def held_curve_years(data_dir: Path, *, first_year: int = FIRST_CURVE_YEAR) -> dict[str, set[int]]:
     """Which years of which feed a previous run already published.
 
     A year of rates that is over cannot change, and each of these files is
@@ -345,7 +370,7 @@ def held_curve_years(data_dir: Path) -> dict[str, set[int]]:
 
     this_year = datetime.now(UTC).year
     held: dict[str, set[int]] = {}
-    for feed, (table, curve_type, column) in FEED_TABLES.items():
+    for feed, (table, curve_type, _column) in FEED_TABLES.items():
         try:
             rows = read_table(data_dir, table).collect()
         except (FileNotFoundError, pl.exceptions.ComputeError):
@@ -354,8 +379,14 @@ def held_curve_years(data_dir: Path) -> dict[str, set[int]]:
             continue
         if curve_type is not None:
             rows = rows.filter(pl.col("curve_type") == curve_type)
-        years = set(rows[column].dt.year().unique().to_list())
-        held[feed] = {year for year in years if year < this_year}
+        if rows.is_empty():
+            continue
+        # Every past year is settled, not only the ones that produced rows.
+        # The real curve starts in 2003 and the years before it return a file
+        # with nothing but a header; treating those as unread meant fetching
+        # thirteen empty years on every run, forever, to be told again that
+        # they are empty.
+        held[feed] = set(range(first_year, this_year))
     return held
 
 
@@ -531,9 +562,14 @@ def held_results(data_dir: Path) -> dict[str, dict]:
     every scheduled run would refetch six thousand artifacts -- forty minutes
     of somebody else's bandwidth to rediscover numbers that cannot move.
 
-    Only rows whose instant came from an artifact are reused. A row dated to
-    the end of its day is one whose artifact was never found, and the next run
-    should look again in case the Treasury has since published one.
+    What decides reuse is `raw_sha256`, not the precision. A row carrying a
+    digest is one whose artifact was fetched and read, and that is settled
+    whether or not the artifact turned out to state a release time -- 1 357 of
+    them do not, and an earlier version of this function kept refetching every
+    one of them forever, looking for a line that was never going to appear.
+
+    A row with no digest is one whose artifact was never found. Those are
+    looked for again, in case the Treasury has since published one.
     """
 
     try:
@@ -545,7 +581,7 @@ def held_results(data_dir: Path) -> dict[str, dict]:
     return {
         row["auction_id"]: row
         for row in rows.iter_rows(named=True)
-        if row["knowledge_time_precision"] == "minute"
+        if row["raw_sha256"]
     }
 
 
@@ -572,6 +608,7 @@ async def build(
         if limit:
             auctions = auctions[-limit:]
         held = held_results(Path(settings.data_dir)) if reuse and not refresh else {}
+        held_times = held_closing_times(Path(settings.data_dir)) if held else {}
         LOGGER.info(
             "%d auctions in the query, %d results already held", len(auctions), len(held)
         )
@@ -589,6 +626,11 @@ async def build(
                 calendar.append(calendar_row(auction, artifact=query, times=times))
                 announcement = announcement_row(auction, artifact=query, times=times)
                 if announcement is not None:
+                    # Carry forward the two fields that came from the artifact
+                    # and are not in the query.
+                    for field, value in (held_times.get(auction.auction_id) or {}).items():
+                        if announcement.get(field) is None:
+                            announcement[field] = value
                     announcements.append(announcement)
                 results.append(known)
                 reused += 1
@@ -609,7 +651,11 @@ async def build(
                 LOGGER.info("%d/%d auctions read", index, len(auctions))
         LOGGER.info("%d results reused, %d read from source", reused, len(results) - reused)
 
-        held_years = held_curve_years(Path(settings.data_dir)) if reuse and not refresh else {}
+        held_years = (
+            held_curve_years(Path(settings.data_dir), first_year=settings.first_curve_year)
+            if reuse and not refresh
+            else {}
+        )
         curves = await fetch_curves(
             fetcher, first_year=settings.first_curve_year, skip=held_years
         )
@@ -722,6 +768,7 @@ __all__ = [
     "calendar_row",
     "curve_features",
     "curve_rows",
+    "held_closing_times",
     "held_curve_rows",
     "held_curve_years",
     "held_results",

tests/test_auctions.py

@@ -162,11 +162,14 @@ def test_an_auction_with_no_date_is_not_an_auction():
     assert isinstance(Auction(auction_id="x", row={}), Auction)
 
 
-def test_only_results_dated_from_an_artifact_are_reused(tmp_path):
-    """A finished auction's artifact cannot change; a missing one might appear.
-
-    Reusing a row that was dated to the end of its day would mean never looking
-    again for an artifact the Treasury had since published.
+def test_a_row_is_reused_when_its_artifact_was_read(tmp_path):
+    """The digest decides, not the precision.
+
+    1 357 artifacts exist and state no release time. Their rows are dated to
+    the end of the auction day, and an earlier version of this rule refetched
+    every one of them on every run, looking for a line that was never going to
+    appear. A row with no digest is one whose artifact was never found, and
+    that one is worth looking for again.
     """
 
     import polars as pl
@@ -176,17 +179,55 @@ def test_only_results_dated_from_an_artifact_are_reused(tmp_path):
     from recipe.store import write_table
 
     rows = [
-        {"result_id": "A", "auction_id": "UST_2020-01-02_X", "knowledge_time_precision": "minute"},
-        {"result_id": "B", "auction_id": "UST_1998-07-27_Y", "knowledge_time_precision": "date"},
+        # Artifact read, stated a release minute.
+        {"result_id": "A", "auction_id": "UST_2020-01-02_X",
+         "knowledge_time_precision": "minute", "raw_sha256": "aaa"},
+        # Artifact read, stated no release time. Settled all the same.
+        {"result_id": "B", "auction_id": "UST_2009-03-05_Y",
+         "knowledge_time_precision": "date", "raw_sha256": "bbb"},
+        # No artifact was ever found. Look again.
+        {"result_id": "C", "auction_id": "UST_1998-07-27_Z",
+         "knowledge_time_precision": "date", "raw_sha256": None},
     ]
     data_dir = tmp_path / "data"
     write_table(data_dir, "auction_results",
                 align(pl.DataFrame(rows, infer_schema_length=None), RESULT_SCHEMA))
     held = held_results(data_dir)
-    assert set(held) == {"UST_2020-01-02_X"}
+    assert set(held) == {"UST_2020-01-02_X", "UST_2009-03-05_Y"}
 
 
 def test_nothing_is_held_when_there_is_no_published_tree(tmp_path):
     from recipe.pipeline import held_results
 
     assert held_results(tmp_path / "nothing") == {}
+
+
+def test_closing_times_that_only_the_artifact_had_survive_a_reuse(tmp_path):
+    """Five auctions of 11 114 state a noncompetitive close only in the artifact.
+
+    Rebuilding their announcement from the query alone would silently drop a
+    field the dataset already had, and the incremental build would stop
+    matching the full one.
+    """
+
+    import polars as pl
+
+    from recipe.pipeline import held_closing_times
+    from recipe.schema import ANNOUNCEMENT_SCHEMA, align
+    from recipe.store import write_table
+
+    rows = [
+        {
+            "announcement_id": "UST_2020-07-13_X_ANN",
+            "auction_id": "UST_2020-07-13_X",
+            "competitive_close_at": datetime(2020, 7, 13, 15, 30, tzinfo=UTC),
+            "noncompetitive_close_at": datetime(2020, 7, 10, 15, 0, tzinfo=UTC),
+        }
+    ]
+    data_dir = tmp_path / "data"
+    write_table(data_dir, "auction_announcements",
+                align(pl.DataFrame(rows, infer_schema_length=None), ANNOUNCEMENT_SCHEMA))
+    held = held_closing_times(data_dir)
+    assert held["UST_2020-07-13_X"]["noncompetitive_close_at"] == datetime(
+        2020, 7, 10, 15, 0, tzinfo=UTC
+    )

tests/test_curves.py

@@ -112,8 +112,14 @@ def test_the_methodology_changed_on_the_date_the_treasury_states():
     assert all(len(text) > 100 for text in table["description"])
 
 
-def test_only_finished_years_of_a_feed_are_reused(tmp_path):
-    """A year that is over cannot change. The current year gains a row a day."""
+def test_every_past_year_is_settled_including_the_empty_ones(tmp_path):
+    """A past year with no rows is a year the Treasury published nothing for.
+
+    The real curve begins in 2003 and the years before it return a file with
+    nothing but a header. Holding only the years that produced rows meant
+    refetching those empty years on every run, forever, to be told again that
+    they are empty. The current year is never held: it gains a row a day.
+    """
 
     from datetime import UTC, datetime
 
@@ -127,13 +133,13 @@ def test_only_finished_years_of_a_feed_are_reused(tmp_path):
     rows = [
         {"curve_id": f"nominal_{year}-06-01_10Y", "curve_date": date(year, 6, 1),
          "curve_type": "nominal", "maturity": "10Y", "yield_pct": 4.0}
-        for year in (this_year - 2, this_year - 1, this_year)
+        for year in (this_year - 2, this_year)
     ]
     data_dir = tmp_path / "data"
     write_table(data_dir, "nominal_yield_curve",
                 align(pl.DataFrame(rows, infer_schema_length=None), CURVE_SCHEMA))
-    held = held_curve_years(data_dir)
-    assert held["nominal"] == {this_year - 2, this_year - 1}
+    held = held_curve_years(data_dir, first_year=this_year - 5)
+    assert held["nominal"] == set(range(this_year - 5, this_year))
     assert this_year not in held["nominal"]
 
 
@@ -155,9 +161,12 @@ def test_two_feeds_sharing_a_table_are_tracked_apart(tmp_path):
     data_dir = tmp_path / "data"
     write_table(data_dir, "nominal_yield_curve",
                 align(pl.DataFrame(rows, infer_schema_length=None), CURVE_SCHEMA))
-    held = held_curve_years(data_dir)
-    assert held["nominal"] == {2010}
-    assert held["long_term_composite"] == {2011}
+    held = held_curve_years(data_dir, first_year=2010)
+    # Both feeds have published rows, so both hold every past year.
+    assert 2010 in held["nominal"]
+    assert 2011 in held["long_term_composite"]
+    # A feed with no rows at all holds nothing.
+    assert "real" not in held
 
 
 def test_nothing_is_held_without_a_published_tree(tmp_path):