samebytes 0.2.0

Three additions. Nothing that v0.1.0 printed, hashed, probed or installed
changes: every v0.1.0 test passes unchanged, the probe set and its grids are
the same, and parity without --lock prints byte-identical output.

• frames.shared_hash(a, b, drop) and frames.shared_columns. A release
  that adds an output column cannot be checked by hashing the whole frame
  against the build before it. economicspace once reported four cells as
  different on a configuration identical on 142 of 142 shared columns.
  shared_hash hashes both frames over the columns both carry, in A's order,
  and returns the column list so the set can be named. column_diff now takes
  its shared set from shared_columns, with the same result.
• provenance, moved from economicspace's verify.py:
  • git_state(repo, paths) gives the commit, the branch, and the tracked
    files under paths whose content differs from HEAD (git diff, never
    git status).
  • environment(packages) gives the interpreter, system, machine, host and
    each package's version.
  • "Could not tell" is None and never reads as clean.
  • A module already imported reports its own __version__; otherwise
    importlib.metadata answers without importing the package.
• parity --lock FILE (repeatable) compares each installed distribution
  with its name==version pin.
  • Names are normalised as PEP 503 does, and versions are compared as exact
    strings.
  • Every other line (a range, a URL, -r, an environment marker) is printed
    NOT CHECKED and fails the lock check.
  • The exit code is the probes' when they did not all match, else the lock's.
  • The lock is checked even when the reference is missing (exit 2).
  • --lock with --record is an argument error.

Proved: 22 new tests, 90 in all. Twelve defects were planted in a copy (whole B
hashed, provenance kept, B's own order, git status for dirty, no git read as
clean, metadata ignoring the imported module, unreadable lock lines dropped or
passed, names not normalised, tab comments, the lock skipped without a
reference, the lock's exit code ignored), and each failed its test, as do v0.1.0's 16.

Also in this release, documentation only:

• README: a Develop section (running the tests, what they need, how a release
  is made), a Licence section, and the --no-config / --require-config
  flags of hooks.
• Docstrings in tree, frames and parity say plainly what they said with
  filler, and a few paragraphs broken by the move out of economicspace are
  rewrapped.
• pyproject.toml: a plainer description, two keywords, Changelog and Issues
  links.

samebytes 0.1.0

First release: economicspace's tree_check.py, platform_check.py and
.githooks/ (as of economicspace 087aef5), with the comparison functions of
its verify.py, made into a package. The commits that built the three moved
files are replayed at the start of this repository's history, limited to
those files; frames.py starts from a verbatim AST extraction so that its
first diff is the whole of what changed.

What changed in moving it:

• The repository is an argument. tree_check.py checked the directory it
  sat in; installed, that is site-packages, so verify_tree(repo) and
  assert_tree(repo, label) take it and nothing defaults it except the
  command line (the current directory).
• parity takes its reference path and its advice. The reference file and
  the four project phrases the report prints (HINTS: the re-baseline
  command, the record command, where the dtype traps and the hashes are
  written down) are arguments. The probes and their argument grids are
  unchanged, so a reference recorded by economicspace stays comparable.
  Proved: with economicspace's hints, report prints byte-identical output to
  platform_check.py on all-match, libm, pandas, hard-defect and combined
  references, and collect() is identical.
• frames is verify.py's comparison as public functions:
  comparable, canonical_csv, frame_hash, column_diff, blanks,
  read_exact, float_hash. The provenance columns are a drop= argument.
  Proved against verify.py on 108 baseline cells: the same frame, hash,
  canonical CSV and column diff for every one.
• One behavioural difference: blanks accepts pd.NA. verify.py's
  version raised "boolean value of NA is ambiguous" on the nullable string
  dtype; it never met one. On every input verify.py accepted, the results are
  the same.
• read_exact refuses a float_precision argument rather than obeying
  it, and passes every other read_csv argument through.
• hooks: install copies the sh hooks into a repository and sets
  core.hooksPath .githooks; check compares the copies byte for byte and
  asks git for the eol=lf pin and the executable bit. The hook's header now
  names samebytes instead of economicspace's CLAUDE.md.
• python -m samebytes tree | parity | hooks install/check.

Tests: the stale same-length read git cannot see, the absent file, CRLF
through .gitattributes, the skip that must not pass; each frames rule; the
parity buckets and the pinned probe grids; every hooks finding and the hook
itself leaving a real edit alone. Sixteen of the bugs behind them are planted
in a copy of the package and shown to fail their test.