[melhzy]

Fixes from a code review of the Python and R packages, one commit per step.

Changes

1. R seeds now match Python. R took the first 7 hex digits of the MD5 digest, so the same string gave a different seed in each language, contrary to the paper. seed_number is now the MD5 digest of the UTF-8 input modulo 2^32, as in Python. In the same commit:
   • SeedHashGenerator$new() defaults to 0 to 2^31−1. It used −1e9 to 1e9, which disagreed with create_seedhash(), Python and the tests; with that default, create_seedhash() always failed in generate_seeds().
   • generate_seeds() returns integers, no longer overflows to NA near 2^31−1, and works on the full integer range.
   • generate_seeds() pins the RNG kind and restores the caller's .Random.seed.
   • Fields are read-only.
   • New set_seed(), because set.seed(seed_number) fails when the seed is above 2^31−1.
2. R tests can now fail CI. The old script printed FAILED but always exited 0, and two of its tests were failing. The tests are now a testthat suite, which CI installs and runs.
3. Python: infinite loop on seed collisions. A repeated seed value could make itself its own ancestor, so SeedExperimentManager.add_experiment_result() never returned.
4. Python packaging.
   • Requires Python 3.8, since typing.Literal broke the import on 3.7.
   • The extras moved to pyproject.toml. setuptools ignored the ones in setup.py, so pip install seedhash[torch] installed nothing extra.
   • Requires torch 1.11, which warn_only=True needs.
5. Python correctness.
   • Binary classification metrics assumed labels 0 and 1.
   • Multi-class F1 now averages per-class F1.
   • (n,) against (n, 1) inputs no longer broadcast silently.
   • The silhouette score handles singleton clusters and identical points correctly.
   • The RL improvement_rate had the wrong sign for negative rewards.
   • Sampling ranges that are too narrow now raise a clear error.
   • generate_seeds() no longer resets the global random state.
   • A broken TensorFlow install no longer breaks import seedhash.
   • The import-time pandas warning is gone.
6. Docs. Corrected the cross-language claim in paper.md and the porting guide.
7. Instructions. Updated the R range and R version docs, and fixed test commands that pointed at a tests/ folder that doesn't exist.

Breaking changes

• R (not yet on CRAN):
  • seed_number and all generate_seeds() output change.
  • The default range of SeedHashGenerator$new() changes.
  • Fields are read-only.
  • Requires R ≥ 3.6.0.
• Python:
  • Requires Python 3.8 or later.
  • Seeds and sampling output for valid inputs are unchanged; I checked 54 cases recorded before the changes.
  • Metric values change where they were wrong.

Testing

• Python: 33 tests pass locally. 15 of them are new in test_correctness.py, and 12 of those fail on the old code.
• R: 59 testthat expectations pass, and R CMD check reports Status OK (it reported 1 NOTE before).
• Seed values for the same strings are pinned in both test suites, so a cross-language mismatch would fail CI.

🤖 Generated with Claude Code

[Copilot]

Copilot review overview

🔵 Needs a closer look

It is a broad, intentionally breaking cross-language change touching seed derivation, RNG-state handling, and numerical metrics whose R test suite could not be executed in this environment, so final human verification is warranted.

Review effort: Balanced
Findings: None

What changed in this PR

This PR applies a batch of correctness, packaging, and documentation fixes uncovered during a code review of the dual-language seedhash package (deterministic MD5-based seed generation). Its central goal is to make the R implementation produce the same seed_number as Python (MD5 digest of the UTF-8 input modulo 2³²), fix several latent Python bugs (an infinite loop on seed collisions, incorrect ML metrics, global-state side effects, fragile optional imports), modernize packaging/CI, and bring the many docs in line. The changes are intentionally breaking (R seed values, R default range, and several Python metric values change) but the package is pre-CRAN and the Python seed/sampling outputs for valid inputs are preserved.

Changes:

• R: rewrite seed derivation to match Python, add read-only fields, set_seed(), overflow-safe generate_seeds() that restores caller RNG state, and a real testthat suite wired into CI.
• Python: fix the add_experiment_result() infinite loop on repeated seeds, correct binary/multi-class/silhouette/RL metrics, add shape validation, stop mutating global random state, and harden optional framework imports.
• Packaging/docs: require Python 3.8 / torch 1.11 / R 3.6.0, move extras to pyproject.toml, and update READMEs, the porting guide, and paper.md.

File	Description
R/​R/​seedhash.R	Core rewrite: Python-matching seed, read-only active bindings, set_seed(), RNG-state-preserving generate_seeds()
R/​tests/​testthat/​test-seedhash.R, testthat.R	New testthat suite with pinned cross-language values
R/​tests/​test_seedhash.R	Removed old always-exit-0 script
R/​man/​*.Rd, R/​NAMESPACE, R/​DESCRIPTION, R/​NEWS.md	Regenerated docs, imports, deps (R ≥ 3.6.0, testthat)
R/​*.md, PYTHON_TO_R_GUIDE.md, GITHUB_INSTALL.md	Updated cross-language and integer-range guidance
Python/​seedhash/​experiment.py	Sampling validation, hierarchy cycle fix, metric corrections, dict copies
Python/​seedhash/​core.py	Robust optional imports, private RNG in generate_seeds, TF/torch determinism tweaks
Python/​setup.py, pyproject.toml, requirements.txt	Python 3.8+, torch 1.11+, extras in pyproject
test_correctness.py	New regression tests (seeds, imports, metrics, hierarchy)
.github/​workflows/​tests.yml	Install R package, run testthat, add correctness tests
README.md, Python/​README.md, paper.md, CONTRIBUTING.md, JOSS_*.md	Doc/test-command updates

I verified the cross-language seed derivation by computation (R's last-8-hex construction equals Python's int(hex,16) % 2³²), reproduced the pinned Python seed/sampling values locally, and traced the metric, sampling-validation, and hierarchy-cycle logic by hand — all consistent. I did not find objective defects meeting the bar for a comment.

Files not reviewed (2)

• R/man/SeedHashGenerator.Rd: Generated file
• R/man/create_seedhash.Rd: Generated file

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.