diff --git a/CHANGELOG.md b/CHANGELOG.md
index ea78dd23..896d687c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -14,7 +14,7 @@ installed, `1.0-alpha`, and `1.0-alpha2`), so a single
 notes in this file describe `default_version` as pinned at an earlier version, each
 true until the next version shipped.
 
-## [Unreleased]
+## [1.0-alpha3] - 2026-09-02
 
 ### Added
 
@@ -224,12 +224,15 @@ true until the next version shipped.
   now raise `22023`.
 
   Retiring a group leaves its old row numbers in the visibility map, so an
-  index-only scan answers from the index for a row group that is gone. `expire`
-  cleared them; `pgcolumnar.recluster()` and the partial-group rewrite behind
-  `pgcolumnar.compact_rewrite()` did not, and both renumber live rows through
-  the same retire. All three clear now, and as of #878 all three are held by
-  tests; before it, only expire's clear was. The rule is that visibility-map bits go
-  wherever row numbers are reassigned, not only where rows expire.
+  index-only scan answers from the index for a row group that is gone. **None of
+  the three paths cleared them.** `expire`'s absence is the one that was
+  reported; `pgcolumnar.recluster()` and the partial-group rewrite behind
+  `pgcolumnar.compact_rewrite()` renumber live rows through the same retire and
+  had the same gap, with no report against either.
+
+  All three clear now. As of #878 all three are held by tests; before it, only
+  expire's clear was. The rule is that visibility-map bits go wherever row
+  numbers are reassigned, not only where rows expire.
 
   `docs/sql-reference.md` gains the accepted range for `ttl_interval` and the
   `NULL` rule, which is stronger than the straddling behaviour the page already
diff --git a/RELEASE_NOTES_1.0-alpha3.md b/RELEASE_NOTES_1.0-alpha3.md
new file mode 100644
index 00000000..be9ff465
--- /dev/null
+++ b/RELEASE_NOTES_1.0-alpha3.md
@@ -0,0 +1,163 @@
+# pgColumnar 1.0-alpha3 release notes
+
+Release date: 2026-09-02
+Previous release: 1.0-alpha2 (2026-08-18)
+
+pgColumnar is a columnar table access method for PostgreSQL. This is the third
+alpha. Its theme is retention and skipping. Rows can now expire on a declared
+interval. A bulk load can refuse work it has already done. Three more predicate
+shapes prune whole chunk groups. The on-disk native format, PGCN v1, is
+unchanged. Existing tables are read and written as before.
+
+This release requires one upgrade command. See "Upgrading" at the end.
+
+## Highlights
+
+- **Retention.** `pgcolumnar.expire` drops row groups whose rows are all older
+  than an interval you declare on the table. It works on whole row groups, so it
+  reclaims space without rewriting live data.
+- **Bulk loads can refuse a repeat.** `pgcolumnar.parallel_copy` records a
+  fingerprint of each load. A load it has already taken is refused rather than
+  duplicated.
+- **Three more predicate shapes prune chunk groups.**
+  `date_trunc(unit, ts)` now drives skipping in both its range and its equality
+  form, and so do `IN (...)` and `= ANY(array)`.
+- **The scan tells the planner what order it is in.** A sorted rewrite leaves an
+  ordering behind. The scan now reports it, so the planner can skip a sort.
+- **The vectorized aggregate takes a wider range of queries**, including a
+  target list that itself contains aggregates.
+
+## Retention
+
+`pgcolumnar.set_options` takes `ttl_column` and `ttl_interval`.
+`pgcolumnar.expire` then retires every row group whose maximum value in that
+column is older than the cutoff. A group is dropped whole. Rows inside the
+retention window are never touched.
+
+The interval must be positive. A negative interval would put the cutoff in the
+future, which would retire groups whose rows are still current.
+
+## Bulk ingest
+
+`pgcolumnar.parallel_copy` records a fingerprint for each completed load in
+`pgcolumnar.load_fingerprint`. Re-running the same load is refused. This makes a
+retried ingest safe to repeat after a failure, without a manual check for
+partial work.
+
+## Skipping and the planner
+
+A predicate on `date_trunc(unit, ts)` now prunes chunk groups. Both the range
+form and the equality form work. `IN (...)` and `= ANY(array)` prune too.
+
+Skip predicates are evaluated most-selective-first, so a group that can be ruled
+out cheaply is ruled out first.
+
+The cost model and the zone-map sample now read the row-group geometry a table
+was **written** with, rather than the current setting. Changing a setting no
+longer reprices existing data.
+
+## Maintenance and reporting
+
+`pgcolumnar.vacuum_sorted()` self-gates. When the relation is already sorted on
+the requested key, it does nothing rather than rewriting the table.
+
+`pgcolumnar.sort_status` reports `sorted_kind`, so a reader can tell which kind
+of ordering a table carries.
+
+## Correctness fixes
+
+Two defects in this list were silent: the operation looked correct and the data
+was wrong. Both were found in the review before this release and both were
+reproduced before being changed.
+
+- **A projection created mid-transaction missed every write that followed it**
+  (#875). A write before `pgcolumnar.add_projection()` in the same transaction
+  left the new projection empty of everything written after it. Measured: 116
+  rows in the base table, 105 in the projection, with no error raised. A
+  covering projection scan then answered as though those rows did not exist. The
+  same defect in `pgcolumnar.drop_projection()` left rows in a projection
+  storage whose catalog rows were already deleted.
+- **An Arrow import ignored the width, sign and scale the file declared**
+  (#881). The importer decoded with the target column's parameters instead of
+  the file's. A `uint64` value above 2^63 was stored as a negative number. An
+  `int64` file read into an `int` column returned `1,0,2,0` for `1,2,3,4`. A
+  `decimal(10,2)` value of 1.25 was stored as 0.0125. A
+  `fixed_size_binary(32)` was read as its first 16 bytes. All four imported
+  without an error and the wrong values were persisted. They are refused now.
+
+  The `1,0,2,0` is worth recognising if you imported integers. The reader took
+  its stride from the target column, so four-byte reads walked an
+  eight-byte-per-value buffer. Every second read landed on the high half of a
+  small positive number, which is zero. The signature is a real value alternating
+  with a zero, not a column of ascending garbage.
+- Index entries for live rows are no longer destroyed (#838).
+- A scrollable cursor no longer answers a backward fetch with forward rows
+  (#842).
+- `sum(bigint)` and `avg(bigint)` no longer accumulate across a rescan (#840).
+- `date_trunc(unit, ts) = 'infinity'` returns the matching row again (#836).
+- An encoded NUL no longer defeats the Iceberg traversal guard (#844).
+- `pgcolumnar.set_options` no longer writes past three stack arrays.
+- Deleted rows no longer count toward the planner's row estimate.
+- A custom scan no longer hides the children of an `INHERITS` parent.
+- `TRUNCATE` retires the old storage's catalog rows, and a transaction that
+  writes, truncates and writes again keeps the rows it should.
+- `ALTER TABLE ... SET ACCESS METHOD heap` drops the catalogs keyed to the
+  relation.
+- A parallel export refuses a destination too long to hold the names it
+  generates.
+- `pgcolumnar.compact_rewrite` and `pgcolumnar.maintenance_due` reject `NaN`
+  thresholds.
+
+## Known issues
+
+- **A rewrite makes a projection read as absent** (#876). `TRUNCATE`, vacuum and
+  recluster mint a new storage id, and the projection rows keep the old one. The
+  projection is still declared and its data is intact.
+  `pgcolumnar.rebuild_projections()` re-records them. The error message names
+  that function.
+- **Two visibility-map clears have tests but no verdict** (#877). The clears on
+  the recluster and partial-rewrite paths gained coverage in this release.
+  Whether a defect sits behind them is not known, and there is no reported
+  symptom.
+
+  The rule they implement is not speculative. Its third application, on
+  `pgcolumnar.expire`, was a real defect: an index-only scan answered from the
+  index for a row group that had been retired. That one is fixed. These two are
+  the same rule on two other paths, with no demonstrated symptom on either.
+
+  Of the two issues in this section, **#876 can affect you today** and has a
+  recovery command. #877 has no known user-visible effect and is listed so the
+  state is on the record, not because there is something to act on.
+
+## Upgrading
+
+Install this build, then run the following in every database that has the
+extension:
+
+```sql
+ALTER EXTENSION pgcolumnar UPDATE;
+```
+
+This is required. The upgrade adds two columns to `pgcolumnar.options` for
+retention and creates the `pgcolumnar.load_fingerprint` table. It creates
+`pgcolumnar.expire(regclass)`, which is the entry point for the retention
+feature above. It replaces four more function definitions:
+`pgcolumnar.parallel_copy` and `pgcolumnar.set_options` are dropped and
+recreated at a new signature, and `pgcolumnar.maintenance_due` and
+`pgcolumnar.sort_status` change in place. No table data is converted and no SQL
+you write changes.
+
+See `docs/installation.md` for the commands, including how to list the databases
+that need the update.
+
+## Scope and limitations
+
+- This is an alpha. Interfaces may change before 1.0.
+- Retention drops whole row groups. A group is retired only when every row in it
+  is outside the retention window.
+- On PGXN this release is `1.0.0-alpha.3`, while `CREATE EXTENSION` reports
+  `1.0-alpha3`. PGXN requires a semantic version, which needs three integer
+  components. The extension's own version has two. The two names refer to the
+  same release.
+
+The complete, itemized list of changes is in `CHANGELOG.md`.
diff --git a/design/RELEASE_PLAN_1.0.md b/design/RELEASE_PLAN_1.0.md
index c830fa4b..fe94a677 100644
--- a/design/RELEASE_PLAN_1.0.md
+++ b/design/RELEASE_PLAN_1.0.md
@@ -27,7 +27,7 @@ both change the bytes on disk, so both belong in an alpha.
 | `v1.0-dev` | yes | the first published marker |
 | `v1.0-alpha` | 2026-08-04 | native format, scan, maintenance, Arrow and Parquet |
 | `v1.0-alpha2` | 2026-08-18 | object storage, Apache Iceberg, parallel bulk ingest and export |
-| `1.0-alpha3` | **not tagged** | retention, `parallel_copy` deduplication, `sort_status` reporting `sorted_kind` |
+| `1.0-alpha3` | 2026-09-02 | retention, `parallel_copy` deduplication, `sort_status` reporting `sorted_kind` |
 
 The observed cadence is 14 days, from one tag to the next. The dates below keep
 it. They are a cadence, not a commitment, and each is the date the tag is cut