[jerome-white]

Closes #13

What this does

README (restructured; the threads and reproducibility section is unchanged):

• Method summary, now stating that the p-value has no +1 correction, so 0 means p < 1/B. Status line updated (pre-release, not registered).
• Installing: Pkg.add(url=...); I ran it in a clean environment and it installs and runs. Julia 1.9 for the library, dependencies Random and Statistics only.
• Using the library: the arguments, the result fields and order, and that the result is a Tables.jl table (checked in Test the Tables.jl interface and drop the unused Tables dependency #29).
• Running the CLI: input format with an example, output columns, and an options table (--B, --seed, --missing); Julia 1.11 for the CLI.
• Missing scores: why a NaN poisons everything, what --missing drop does and when it is sound, and why nanmean is deliberately not offered (decisions from Warn when scores contain NaN #22/Reject NaN in the library; add --missing to the CLI #23).
• Known limitation: floating-point ties (p-values depend on floating-point rounding when a permuted range equals the observed difference #17): until now documented only in the tests. States the effect, the example (0.3935 against 0.4028 for tiny.csv) and that there is deliberately no tolerance.
• How it is checked: the test suite, CI matrix, and the exact and Sakai (Discpower) references; pointer to benchmark/.
• Layout table.

Docstring of randomized_tukey_hsd (the only change to src/, text only): lists every error it throws (B, fewer than two systems, NaN, DimensionMismatch), says what p = 0 means, and adds a short runnable example. The ragged line wrapping from the rng edits is fixed.

Decisions

• No Documenter.jl. The issue says "only if it proves worthwhile". There is one exported function, and the README plus its docstring cover it; Documenter would add a docs environment, a build in CI and a hosting step for one page. Easy to revisit if the API grows.
• Not repeated here: the performance numbers and the design records live in benchmark/README.md and the issues (Perf/safety: thread scheduling and per-thread state #6, Fixed-seed results independent of thread count #24, Make the RNG an explicit parameter #26, Decide how the chunk size is exposed #32); the README links the benchmark directory and states only what a user has to know.

Verification

• The docstring example runs (pvalue 0.0048 for A against C at B = 10 000, seed 1).
• The Pkg.add(url=...) install and a call, in a clean environment.
• The CLI usage and output columns, run on test/data/tiny.csv.
• ] test passes (96 pass + 1 deliberately broken); 80 columns checked.
• Not checked: the rendered Markdown on GitHub; the anchor links (#missing-scores, #threads-and-reproducibility) rely on GitHub's usual heading ids.

🤖 Generated with Claude Code