[pranav2012]

Summary

Ports abtestify to Go, Rust, Python, Java (which also serves Kotlin) and Swift. Every port produces byte-for-byte the same assignments as the TypeScript implementation, and one release publishes all of them with the same version. The TypeScript package stays where it is; the npm tarball is unchanged (still only dist).

Language	Location	API	Registry
Go	go/	GetBucket(Params) (Result, error). Optional fields are pointers (abtestify.Float64, abtestify.String); returns *ValidationError{Field}	tag go/vX.Y.Z
Rust	rust/	get_bucket(&Params) -> Result<Assignment, Error>, with a Params::new(..) builder and Default. Zero deps (own SHA-256), MSRV 1.60	crates.io
Python	python/	get_ab_experiment_bucket(experiment_name, unique_id, *, treatment_split=50, ...) -> Assignment (frozen dataclass). Uses hashlib, typed with py.typed, 3.9+	PyPI (trusted publishing)
Java/Kotlin	java/	ABTestify.getBucket(Params) / getBucket(name, id), Params.builder(..), an immutable Assignment, and InvalidParamsException extends IllegalArgumentException. Java 8 bytecode	Maven Central io.github.pranav2012:abtestify
Swift	Package.swift (root) + swift/	ABTestify.bucket(experimentName:uniqueId:...) throws -> Assignment. Zero deps, no CryptoKit, builds on Linux	the existing plain x.y.z release tags

How parity is verified

test-vectors/generate.mjs writes test-vectors/vectors.json from the TypeScript implementation. Assignments come from the built dist/index.js; the internal bucketFromTop/bucketOf from src/bucket.ts are compiled with esbuild. The generator uses a seeded PRNG, so its output is deterministic.

• 2,859 hashing cases: ASCII, Japanese, emoji, ZWJ sequences, combining marks (NFC vs NFD), NUL and astral-plane characters, IDs up to 10,000 chars across the SHA-256 padding boundaries, splits and exposures of 0, 100 and fractions like 33.3, thresholds set exactly on each user's bucket, legacyHashing both ways, and custom variant names (including the empty string).
• 4,919 bucket mapping cases (top64Hex to bucket): 0, 1, 2^53±1, 2^63±1, 2^64−1, ties at every exponent from 2^53 to 2^63 (e.g. 2^54+2, 2^63+1536), and values at and ±1/±2/±2048 around the exact float boundary of ~400 buckets, plus the naive ceil(k·2^64/10000) boundaries. I checked that a truncating conversion fails 1,658 of these, ties-away fails 209, and exact integer math fails 1,667.
• 26 error cases: negative, >100, ±1e308, NaN and ±Infinity (as strings) for both percentages, empty IDs, and the order of the checks.

bucket.ts now exposes bucketFromTop (a pure refactor). A new jest test checks the TypeScript implementation against the vectors, and CI regenerates the file and fails if it differs from the committed copy. Java's unsigned conversion uses ((x >>> 1) | (x & 1)) * 2.0 and is also checked against BigInteger.doubleValue() for every bucket case.

Local test results (macOS arm64)

• TypeScript: yarn test passes: 52 tests, including the vector suite. Lint, prettier and type-check are clean.
• Go: go test ./... passes on arm64, and on amd64 under Rosetta. 386 runs in CI only.
• Rust: cargo test passes (5 unit tests + 1 doctest). Clippy and fmt are clean, cargo package succeeds, and the library builds on Rust 1.60.
• Python: pytest passes on 3.13 (7 tests + 26 subtests). mypy --strict is clean and the wheel/sdist build. 3.9 to 3.12 run in CI only.
• Java: ./gradlew test passes on JDK 17 (bytecode major 52), and publishToMavenLocal produces the jar, sources, javadoc and POM. Signing is not tested locally because there's no gpg here.
• Swift: swift test passes on macOS with Swift 6.3. Linux (swift:6.0 container) runs in CI only.
• Workflows: actionlint is clean, apart from pre-existing shellcheck info notes in the untouched npm version-bump script.

CI and release

• pr-verify.yml runs these jobs in parallel: typescript (with the vectors-fresh check), go (1.18 and stable, plus 386), rust and rust-msrv, python 3.9 to 3.13 plus a package check, java, swift-linux and swift-macos.
• release.yml keeps the npm job and its version logic unchanged, apart from bumping checkout and setup-node to v4. After the GitHub release (whose tag is also the Swift release), these jobs run with the same new_version:
  • tag-go pushes go/v<version>.
  • publish-crates publishes to crates.io.
  • publish-pypi publishes through trusted publishing in environment pypi.
  • publish-maven publishes to Maven Central through the vanniktech plugin with in-memory signing.
• Each registry job skips with a ::warning:: until its credentials exist. PyPI is gated on the variable vars.PYPI_ENABLED == 'true'.

Owner checklist before merging (details in RELEASING.md)

• crates.io: verify your email, create a token with publish-new/publish-update, and add it as the secret CARGO_REGISTRY_TOKEN.
• PyPI: add a pending publisher (project abtestify, owner pranav2012, repo abtestify, workflow release.yml, environment pypi), create the GitHub environment pypi, and set the repo variable PYPI_ENABLED=true.
• Maven Central: create a Central Portal account (signing in with GitHub verifies io.github.pranav2012), generate a user token, and add it as MAVEN_CENTRAL_USERNAME / MAVEN_CENTRAL_PASSWORD.
• GPG: create a key with a passphrase, upload the public key to keyserver.ubuntu.com, and add the armored private key as SIGNING_KEY and the passphrase as SIGNING_PASSWORD.
• Decide the version. package.json is 0.2.0 and the latest release is 0.2.1, so merging as-is releases 0.2.2 everywhere. Set package.json to 0.3.0 first if you want a minor release for the new ports.
• Any secret you skip only skips that registry; npm still releases.