python_extensions / cpython-extensions 1.1.0
Release certification
Date: 2026-08-21

SCOPE
-----
1.1.0 is the performance-evidence, verification, and documentation milestone
for the CPython 3.13 release line. The production transformation modules are
byte-identical to the current 1.0.4 source tree; only
src/python_extensions/_version.py changes inside the runtime package. The new
release therefore preserves the established switch/inline/goto semantic
contracts while adding stronger structural regressions, benchmark evidence,
and release/documentation consistency checks.

SUPPORTED BOUNDARY
------------------
- Implementation: CPython
- Python: >=3.13,<3.14
- Primary certification runtime: CPython 3.13.5
- Runtime dependency: bytecode>=0.17,<0.18
- License: MPL-2.0
- Free-threaded CPython 3.13t: not independently certified
- CPython 3.14+, 3.12-, PyPy/GraalPy: not certified

RUNTIME IDENTITY CHECK
----------------------
Recursive comparison against the 1.0.4 runtime package found exactly one
changed runtime file:

  src/python_extensions/_version.py   1.0.4 -> 1.1.0

Representative unchanged SHA-256 digests:

  switch.py       94b6fdff464c934d7672e5d5bb6e3bfaac0e54d8b84e26e6dfa3aded1ef7588f
  inline.py       0fdb545b2e8e4fcdd5026cfb16b765ca94edc7091e9441c0fd2d8a9073d97472
  goto.py         d50eeda8871ca4857e60d527b72c925b05917a5a7089d11756679247dc668de2
  _core/verify.py 66a6f1b6bd00f9baddabf28aaa0ae02f869b291ce4a1b9e5b562b423fe563906

REGRESSION AND DEVELOPMENT-MODE GATES
-------------------------------------
- python -m pytest -q
  PASS: 378/378
- PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -X dev -W error -m pytest -q
  PASS: 378/378
- python -m compileall -q src tests tools benchmarks/scripts
  PASS

REPRODUCIBLE ARTIFACT GATES
---------------------------
The 1.1.0 source tree was built twice with the repository's setuptools backend
and a fixed SOURCE_DATE_EPOCH. The resulting wheel, canonicalized sdist, and
checksum manifest were byte-for-byte identical across both builds.

- SHA256SUMS verification: PASS
- Exact wheel install + tools/smoke_release.py: PASS
- Exact sdist full regression suite: PASS, 378/378
- Wheel rebuilt from the exact sdist with the same epoch: byte-identical PASS
- Wheel metadata version: 1.1.0
- Wheel CPython classifier: Programming Language :: Python :: Implementation :: CPython

The local sandbox did not include the optional Twine / trove-classifiers
executables used by the hosted release job. The release workflow installs the
declared build tooling before running metadata preflight and `twine check`;
publication remains conditional on those hosted gates.

FULL-SCALE DIFFERENTIAL / STRESS RERUN
--------------------------------------
- Production switch differential/stress harness:
  PASS: 3,680,000 calls
- Adversarial switch harness:
  PASS: 4,703,400 calls
- Strict goto harness:
  PASS: 3,526,500 calls/yields
- Coordinated all-extension harness:
  PASS: 157,200/157,200 calls
- Inline optimizer randomized differential:
  PASS: 1,000,000 rounds
- Inline optimizer threaded differential:
  PASS: 800,000 calls
- Inline unbound-local semantics:
  PASS: 100,000 rounds
- Crash-isolated optimized inline calls:
  PASS: 2,000,000 calls
- Guarded binding harness:
  PASS: 5,110,000 operations

Total explicitly counted operations/calls/yields above: 21,077,100.

1.1.0 SWITCH SCALING EVIDENCE
-----------------------------
Driver:
  benchmarks/scripts/benchmark_switch_scaling_v110.py
Evidence:
  benchmarks/results/BENCHMARK_SWITCH_SCALING_V110.json
Runtime:
  CPython 3.13.5 / Linux x86-64
Method:
  timeit.repeat; uniform forward+reverse successful hits; misses validated but
  excluded from timing; 100,000 target dispatches/sample through 256 routes,
  scaled by route count thereafter with a 50,000-dispatch floor; 7 repeats;
  50 warm-up batches.

Selected median integer results versus generated if/elif:
- 64 routes:    9.94x
- 256 routes:  38.96x
- 512 routes:  71.33x
- 1,024 routes: 142.90x

The 1,024-route string case records 83.86x versus if/elif. The specialized
1,024-route direct-value switch remains 74 bytes of executable bytecode while
the generated integer if/elif baseline reaches 17,420 bytes and match reaches
19,466 bytes. Route-table storage still grows with route count; the bounded
claim applies to the executable dispatch path.

Small-route crossover is retained in the public evidence: at two routes native
branching remains faster. The release therefore does not claim a universal
switch speedup.

INLINE / GOTO INTENDED-WORKLOAD EVIDENCE
----------------------------------------
Driver:
  benchmarks/scripts/benchmark_extension_benefits_v110.py
Evidence:
  benchmarks/results/BENCHMARK_EXTENSION_BENEFITS_V110.json

Recorded medians:
- Frozen small-helper inlining: 1.34x faster than the ordinary helper call.
- Strict goto explicit 3-state machine: 2.47x faster than explicit state
  dispatch.
- Naturally structured reference remains about 1.1% faster than strict goto,
  which is documented as an important interpretation boundary.

DOCUMENTATION / RELEASE CONSISTENCY
-----------------------------------
The 1.1.0 pass audits and synchronizes README, changelog, citation metadata,
security policy, contributor guide, GitHub/PyPI setup, repository metadata,
architecture, compatibility, comprehensive guide, release notes, release
procedure, benchmark guide/evidence, issue template, and release-sensitive
repository hygiene checks.

The source-level goto/label notation credits Entrian's "goto for Python" as an
inspiration while explicitly documenting that cpython-extensions uses an
independent CPython 3.13 lowering and verification implementation.

VERDICT
-------
PASS. The 1.1.0 tree is suitable as the next supported release line, subject
to the normal GitHub Actions reproducible-build, metadata-preflight, GitHub
Release, protected-environment approval, and PyPI Trusted Publishing gates.
