Build and verify Linux arm64 CPython 3.14t wheels

[justrach]

Summary

• build CPython 3.14t Linux aarch64 wheels on a native GitHub ARM64 runner with cibuildwheel
• compile inside manylinux and let auditwheel repair/tag the artifact instead of manually claiming compatibility
• install the repaired wheel in an isolated environment, require the Zig backend, start TurboAPI, and make a real HTTP request
• add an executable Apple container Linux arm64 recipe with pinned Python/Zig/runtime inputs and a host-to-guest HTTP smoke test
• restrict PyPI publication to validated v* tag runs

Verification performed in this session

• Apple container 0.11.0 build and run on Apple Silicon, Linux aarch64 guest
• CPython 3.14.6 free-threaded runtime (Py_GIL_DISABLED=1)
• wheel-installed turbonet.cpython-314t-aarch64-linux-gnu.so loaded from site-packages
• native backend reported zig
• TurboAPI Zig server listened on port 8080 and returned the expected JSON through host port 18080
• negative test confirmed a missing native extension is rejected rather than accepted as simulation mode
• local packaging test produced a cp314-cp314t wheel containing turbonet
• cibuildwheel 4.1.0 selected cp314t-manylinux_aarch64
• release version tests, ruff, Python compilation, shell syntax, git diff --check, and pre-commit tests passed

The GitHub Actions ARM64 job is the source of truth for the repaired manylinux tag and must pass before merge. Publishing the wheel still requires a subsequent versioned release tag.

Closes #188
Closes #189

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Repo admins can enable using credits for code reviews in their settings.

[github-actions]

Performance Regression Report

Runner: ci-pr | Duration: 5s per endpoint | Threads: 4 | Connections: 100 | Workers: 24

Endpoint	req/s	avg latency	p99 latency	threshold	status
GET /health	96,663	0.18ms	1.05ms	80,000	OK
GET /	98,881	0.20ms	1.12ms	80,000	OK
GET /json	94,013	0.19ms	0.87ms	75,000	OK
GET /users/123	96,725	0.20ms	0.73ms	75,000	OK
POST /items	57,385	0.34ms	0.87ms	45,000	OK
GET /status201	86,655	0.20ms	0.67ms	65,000	OK
AVERAGE	88,387			50,000	OK

✅ All endpoints pass regression thresholds

[github-actions]

Performance Regression Report

Runner: ci-pr | Duration: 5s per endpoint | Threads: 4 | Connections: 100 | Workers: 24

Endpoint	req/s	avg latency	p99 latency	threshold	status
GET /health	85,973	0.24ms	0.99ms	80,000	OK
GET /	85,706	0.22ms	1.11ms	80,000	OK
GET /json	80,559	0.22ms	0.92ms	75,000	OK
GET /users/123	81,778	0.23ms	0.86ms	75,000	OK
POST /items	47,151	0.44ms	0.94ms	45,000	OK
GET /status201	77,182	0.25ms	0.79ms	65,000	OK
AVERAGE	76,391			50,000	OK

✅ All endpoints pass regression thresholds

[github-actions]

Performance Regression Report

Runner: ci-pr | Duration: 5s per endpoint | Threads: 4 | Connections: 100 | Workers: 24

Endpoint	req/s	avg latency	p99 latency	threshold	status
GET /health	83,137	0.22ms	1.01ms	80,000	OK
GET /	83,638	0.23ms	1.02ms	80,000	OK
GET /json	80,410	0.44ms	1.46ms	75,000	OK
GET /users/123	82,948	0.25ms	0.69ms	75,000	OK
POST /items	46,789	0.50ms	1.35ms	45,000	OK
GET /status201	75,142	0.24ms	0.77ms	65,000	OK
AVERAGE	75,344			50,000	OK

✅ All endpoints pass regression thresholds

[justrach]

Follow-up verification: glibc portability and POST regression

Linux arm64 wheel

The final artifact from Build & Publish run 29099807825 passed the native ARM64 wheel job. I downloaded and inspected the wheel in this session:

• artifact: turboapi-1.0.31-cp314-cp314t-manylinux_2_28_aarch64.whl
• undefined entropy symbol: getrandom@GLIBC_2.25
• no undefined arc4random_buf
• highest GLIBC symbol version reported by the ELF version table: GLIBC_2.28
• clean manylinux wheel install loaded the extension and served the native HTTP smoke route

POST /items regression

The red gate was a real hot-path regression, not a threshold problem. Commit 2b243d0 installed a new closure on every request-model instance, adding per-request work and creating a model/closure reference cycle. 225b99f replaces that with a route-local compatibility subclass only for DHI models with list fields; the scalar benchmark model is not wrapped. Custom model_dump implementations are not modified.

I reran the unchanged PR quick suite three times at exact head SHA 225b99f54b2b188f8210fca12fb5105cf969c07d using Actions run 29099808178 (job attempts 86385473922, 86385935129, and 86386378208). Command and recorded environment:

BENCH_THREADS=4 \
BENCH_CONNECTIONS=100 \
BENCH_DURATION=5 \
BENCH_RUNNER=ci-pr \
BENCH_VCPUS=2 \
BENCH_OS=ubuntu-latest \
uv run --python 3.14t python benchmarks/bench_regression.py --ci

The runner uses 24 TurboAPI workers and disables TurboAPI response caching and rate limiting. This is the HTTP-only regression suite; no database is involved.

For POST /items, the three recorded attempts produced:

• throughput range: 46,789.37..57,384.77 req/s
• median: 47,151.38 req/s
• p99 latency range: 0.87..1.35 ms
• p99 median: 0.94 ms
• sample count: 3

All three attempts passed the existing 45,000 req/s floor. These are reruns performed in this session and recorded in the cited Actions run; they are not a release benchmark or a cross-framework claim.