0.2.0 — scan, normalize, strip_invisible

Added

• cjkfold scan — every character that breaks a comparison, in one read-only
  pass: the look-alikes, the NFKC family fold() refuses to touch, and the
  invisible family. Each finding says what to do with it — fold, nfkc,
  strip, or manual, where manual means do not touch this without looking.
  It exits 0 by default, because it is a report and not a gate: a gate that fails
  on a load-bearing joiner fails on ordinary emoji. --strict gates anyway,
  --json feeds a pipeline, --classes narrows the report. (c5be88a, 5835dbf)
• normalize() and strip_invisible() — the two explicit writes. The
  first applies the NFKC family with full-width punctuation spared by default; the
  second deletes invisible characters, safe by default, because half of that
  family carries meaning — a family of three is five code points, and deleting the
  joiners leaves three separate people. Neither is exposed as a subcommand: fold -w
  is already an in-place rewrite, and two more destructive commands would be a gun
  pointed at a corpus. (a733ef5)
• Finding, scan_stats() and is_punctuation() in the public API.
  (2bd05e6, 5835dbf)

Changed

• scan's whitelist lookup is now a single set test instead of a dict lookup plus
  four range comparisons: 0.42s → 0.29s over 640,000 characters. (f6c5f4f)

Docs

• README: "The characters that never show up" and "The other two rewrites",
  both carrying output from real runs rather than hand-copied examples; the 中文简述
  follows. The check example also gained the summary line it had been missing —
  the reason it was missing is that it had been copied by hand, which is exactly
  what the new section refuses to do. (e9c4d68, and this release)
• tests/fixtures/ — the three files the roadmap asked for, with check's output
  frozen byte for byte. (1a468e5)

No change to fold(), to check, or to the equivalence table. Text that folded
correctly under 0.1.2 folds identically under 0.2.0.