diff --git a/.github/workflows/tests.yml b/.github/workflows/tests.yml
index adf2660..aff6f1f 100644
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -1,15 +1,40 @@
 # This workflow will install Python dependencies, run tests and lint with a variety of Python versions
 # For more information see: https://docs.github.com/en/actions/automating-builds-and-tests/building-and-testing-python
 
-# TODO:
-# - cache this directory $HOME/.cache/pyensembl/
-# - update coveralls
-# - get a badge for tests passing
-# - download binary dependencies from conda
 name: Tests
 on: [push, pull_request]
 
 jobs:
+  packaging:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        include:
+          - python-version: "3.9"
+            setuptools: "setuptools==77.0.3"
+          - python-version: "3.11"
+            setuptools: "setuptools"
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Install build and test dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install --upgrade build twine wheel ruff pytest pytest-cov '${{ matrix.setuptools }}' -r requirements.txt
+      - name: Build and validate distributions
+        run: |
+          python -m build --no-isolation
+          python -m twine check --strict dist/*
+      - name: Run lint and tests from the source distribution
+        run: |
+          archive_dir=$(mktemp -d)
+          tar -xzf dist/gtfparse-*.tar.gz -C "$archive_dir"
+          cd "$archive_dir"/gtfparse-*
+          ./lint.sh
+          ./test.sh -W error::DeprecationWarning
+
   build:
     runs-on: ubuntu-latest
     strategy:
diff --git a/.gitignore b/.gitignore
index ba74660..65637cb 100644
--- a/.gitignore
+++ b/.gitignore
@@ -39,6 +39,7 @@ htmlcov/
 .coverage
 .coverage.*
 .cache
+.hypothesis/
 nosetests.xml
 coverage.xml
 *,cover
diff --git a/.hypothesis/unicode_data/13.0.0/charmap.json.gz b/.hypothesis/unicode_data/13.0.0/charmap.json.gz
deleted file mode 100644
index 67b0e0b..0000000
Binary files a/.hypothesis/unicode_data/13.0.0/charmap.json.gz and /dev/null differ
diff --git a/MANIFEST.in b/MANIFEST.in
new file mode 100644
index 0000000..ba435b8
--- /dev/null
+++ b/MANIFEST.in
@@ -0,0 +1,3 @@
+include lint.sh test.sh
+recursive-include tests *.py
+recursive-include tests/data *.gtf *.gtf.gz
diff --git a/gtfparse/__init__.py b/gtfparse/__init__.py
index a3dfadd..b0b9c25 100644
--- a/gtfparse/__init__.py
+++ b/gtfparse/__init__.py
@@ -24,7 +24,7 @@
 )
 from .write_gtf import write_gtf
 
-__version__ = "2.9.0"
+__version__ = "2.9.1"
 
 __all__ = [
     "GENCODE_BIOTYPE_ALIASES",
diff --git a/gtfparse/read_gtf.py b/gtfparse/read_gtf.py
index 6b1f12c..eef49ac 100644
--- a/gtfparse/read_gtf.py
+++ b/gtfparse/read_gtf.py
@@ -197,11 +197,11 @@ def parse_gtf_and_expand_attributes(
     filepath_or_buffer, restrict_attribute_columns=None, features=None, *, progress_callback=None
 ):
     """
-    Parse lines into column->values dictionary and then expand
+    Parse a GTF into a Polars DataFrame and then expand
     the 'attribute' column into multiple columns. This expansion happens
     by replacing strings of semi-colon separated key-value values in the
     'attribute' column with one column per distinct key, with a list of
-    values for each row (using None for rows where key didn't occur).
+    values for each row (using empty strings for rows where the key didn't occur).
 
     Parameters
     ----------
@@ -324,7 +324,7 @@ def read_gtf(
     expand_attribute_column : bool
         Replace strings of semi-colon separated key-value values in the
         'attribute' column with one column per distinct key, with a list of
-        values for each row (using None for rows where key didn't occur).
+        values for each row (using empty strings for rows where the key didn't occur).
 
     infer_biotype_column : bool
         Due to the annoying ambiguity of the second GTF column across multiple
@@ -338,7 +338,7 @@ def read_gtf(
 
     column_cast_types : dict, optional
         Dictionary mapping column names to dtypes. Will cast columns to given
-        Polars types.
+        pandas-compatible types.
 
     usecols : list of str or None
         Restrict which columns are loaded to the give set. If None, then
diff --git a/gtfparse/write_gtf.py b/gtfparse/write_gtf.py
index a404740..af7a62e 100644
--- a/gtfparse/write_gtf.py
+++ b/gtfparse/write_gtf.py
@@ -59,8 +59,7 @@ def _attribute_expr(attribute_columns: list[str]) -> polars.Expr:
     empty string (string columns) and cannot distinguish absent from
     present-but-empty, so we treat both null and "" as absent. Values that are
     merely falsy but non-empty -- notably the string "0" -- are written and
-    survive a round trip (this is the regression that the naive ``if value:``
-    check in earlier drafts got wrong).
+    survive a round trip.
     """
     if not attribute_columns:
         return polars.lit("")
diff --git a/pyproject.toml b/pyproject.toml
index 3f10af2..f9c5c1b 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,5 +1,5 @@
 [build-system]
-requires = ["setuptools>=61.0", "wheel"]
+requires = ["setuptools>=77.0.3", "wheel"]
 build-backend = "setuptools.build_meta"
 
 [project]
@@ -7,12 +7,13 @@ name = "gtfparse"
 requires-python = ">=3.9"
 authors = [ {name="Alex Rubinsteyn", email="alex.rubinsteyn@unc.edu" } ]
 description = "Parsing library for extracting data frames of genomic features from GTF files"
+license = "Apache-2.0"
+license-files = ["LICENSE"]
 classifiers = [
     "Development Status :: 4 - Beta",
     "Environment :: Console",
     "Operating System :: OS Independent",
     "Intended Audience :: Science/Research",
-    "License :: OSI Approved :: Apache Software License",
     "Programming Language :: Python",
     "Programming Language :: Python :: 3",
     "Programming Language :: Python :: 3.9",
@@ -71,13 +72,13 @@ select = [
 ]
 ignore = [
     "E501",   # line too long (handled by formatter)
-    "E741",   # ambiguous variable name (pre-existing in codebase)
-    "B006",   # mutable default args (pre-existing; behavior-risky to change)
+    "E741",   # ambiguous variable names
+    "B006",   # mutable default arguments
     "B008",   # do not perform function calls in argument defaults
     "B905",   # zip() without explicit strict
     "SIM108", # use ternary operator instead of if-else
     "UP007",  # use X | Y for type unions (need Python 3.10+)
-    "UP031",  # %-format strings (pre-existing; out of scope for config PR)
+    "UP031",  # %-format strings
 ]
 
 [tool.ruff.lint.per-file-ignores]
diff --git a/tasks/todo.md b/tasks/todo.md
index b5b2c81..7737ed7 100644
--- a/tasks/todo.md
+++ b/tasks/todo.md
@@ -137,3 +137,58 @@ file I/O and output conversion; see the PR validation record for its result.
 
 Release completion is recorded in the PR after CI, merge, and deployment so
 the deployment checkout remains clean.
+
+# Packaging and cleanup: #53, #76, #83
+
+## Specification
+
+The 2.9.0 source distribution contains test modules but omits `tests/__init__.py`,
+`tests/data.py`, and the GTF fixtures. Reproduced six collection errors from an
+extracted release tarball. Add a manifest that ships all test Python modules,
+GTF/gzip fixtures, and lint/test entry scripts in the source distribution.
+Keep tests out of the runtime wheel and do not include generated caches or the
+unused 10 MB parquet artifact.
+
+Use SPDX `Apache-2.0` project license metadata, explicitly include `LICENSE`,
+and raise the setuptools build requirement to 77.0.3, which supports PEP 639.
+Remove the deprecated license classifier. Validate wheel and sdist metadata
+and the embedded license text with both the minimum and current build backend.
+
+Add a packaging CI job on Python 3.9/minimum setuptools and Python 3.11/current
+setuptools. Build with the selected backend, check distributions with twine,
+extract the tarball in a temporary directory, and run the shipped `lint.sh` and
+`test.sh` there. The existing runtime test matrix and coverage aggregation stay
+in place. No unit tests for comment-only edits; real archive execution guards
+the packaging bug.
+
+Complete #76's remaining comment/docstring cleanup and remove the two tracked
+Hypothesis cache files; ignore future Hypothesis caches. Preserve lint rules
+and script behavior. Bump 2.9.0 to 2.9.1; leave the existing README PR #74 alone.
+
+## Plan
+
+- [x] Read issues, repository guidance, package contents, and setuptools docs.
+- [x] Reproduce the release-tarball failure and create a feature branch.
+- [x] Check in with the packaging fix and verification plan.
+- [x] Apply packaging metadata, manifest, CI, and comment/cache changes.
+- [x] Run `./lint.sh` and `./test.sh` with deprecations as errors.
+- [x] Verify both build backends, distribution metadata, extracted-archive
+  lint/tests, and wheel contents.
+- [ ] Open PR, pass CI, merge, run `./deploy.sh` from clean master, and verify
+  the published wheel and source archive.
+
+## Review
+
+`./lint.sh` and all 143 tests pass with deprecations treated as errors in the
+checkout and in extracted source archives built by setuptools 77.0.3 and 84.0.0.
+Both backends build wheels from their source archives successfully, and all
+four distributions pass `twine check --strict`. Verified that source archives
+contain the helper, initializer, GTF fixtures, and lint/test scripts, but no
+Hypothesis/bytecode caches or unused parquet file. Wheels contain no tests.
+Wheel and source metadata both declare `License-Expression: Apache-2.0` and
+`License-File: LICENSE`; included license bytes match the repository file.
+Neither build emits the old license-classifier warning. Workflow YAML parses,
+and git ignores Hypothesis caches at both previously tracked paths.
+
+Release verification will be recorded on the PR after deployment, keeping the
+master checkout clean.
diff --git a/test.sh b/test.sh
index 97556bc..e40eb95 100755
--- a/test.sh
+++ b/test.sh
@@ -1,7 +1,6 @@
 #!/usr/bin/env bash
 # Run the gtfparse test suite with a memory- and CPU-aware pytest-xdist
-# worker count. See ~/code/trufflepig/test.sh for the rationale: running
-# several sibling repos' suites concurrently can fork-bomb the laptop,
+# worker count. Running several test suites concurrently can exhaust resources,
 # so we cap workers at min(cpu_reserve, available_RAM / PER_WORKER_GB).
 # xdist is optional — fall back to serial pytest when it isn't installed.
 #
diff --git a/tests/.hypothesis/unicode_data/13.0.0/charmap.json.gz b/tests/.hypothesis/unicode_data/13.0.0/charmap.json.gz
deleted file mode 100644
index 414a302..0000000
Binary files a/tests/.hypothesis/unicode_data/13.0.0/charmap.json.gz and /dev/null differ
diff --git a/tests/data.py b/tests/data.py
index 23aaacb..bca618e 100644
--- a/tests/data.py
+++ b/tests/data.py
@@ -3,7 +3,6 @@
 
 def data_path(name):
     """
-    Return the absolute path to a file in the varcode/test/data directory.
-    The name specified should be relative to varcode/test/data.
+    Return the absolute path to a fixture relative to tests/data.
     """
     return os.path.join(os.path.dirname(__file__), "data", name)
diff --git a/tests/test_expand_attribute_column_false.py b/tests/test_expand_attribute_column_false.py
index b06a206..ab74485 100644
--- a/tests/test_expand_attribute_column_false.py
+++ b/tests/test_expand_attribute_column_false.py
@@ -1,7 +1,4 @@
-"""Regression tests for #56: read_gtf(expand_attribute_column=False)
-used to raise NameError because the else branch referenced `result_df`
-before it had been assigned.
-"""
+"""Reading a GTF without expanding its raw attribute column."""
 
 import pandas as pd
 
diff --git a/tests/test_gencode_gtf.py b/tests/test_gencode_gtf.py
index 061928a..a344847 100644
--- a/tests/test_gencode_gtf.py
+++ b/tests/test_gencode_gtf.py
@@ -1,4 +1,4 @@
-"""Tests for GENCODE attribute aliases (#63) and version-column casting (#64)."""
+"""Tests for GENCODE attribute aliases and version-column casting."""
 
 import gzip
 import logging
@@ -14,7 +14,7 @@
 
 
 # -----------------------------
-# attribute_aliases (#63)
+# Attribute aliases
 # -----------------------------
 
 
@@ -99,7 +99,7 @@ def test_attribute_aliases_with_empty_dict_is_noop():
 
 
 # -----------------------------
-# cast_version_columns (#64)
+# Version-column casting
 # -----------------------------
 
 
@@ -157,7 +157,7 @@ def test_version_columns_missing_when_attribute_absent():
 
 
 # -----------------------------
-# Interaction between #63 and #64
+# Interaction between aliases and version-column casting
 # -----------------------------