Handling Unicode text correctly

Goal

Process text from any language without corrupting it, and compare or search it in a way users find correct.

Prerequisites

A language with a distinct text type (Python's str) and byte type (bytes), and knowledge of the encodings at each boundary.

Steps

1. Decode input bytes to text as early as possible, with the declared encoding (UTF-8 unless a protocol says otherwise); reject or replace invalid sequences deliberately rather than by accident.
2. Keep all internal processing on text; encode to UTF-8 only when writing to files, sockets or headers.
3. Normalise before comparing or indexing: NFC for storage and display, NFKC when compatibility characters should match (for example, fullwidth digits). Unicode TR15 defines the forms.
4. Treat lengths carefully: byte length for storage limits, code-point length for most APIs, grapheme clusters for what users see. Validate limits in the unit that the limit is about.
5. Case-fold, not lowercase, for case-insensitive comparison of non-ASCII text.
6. Keep identifiers that go into URLs, headers or file names ASCII (transliterate or percent-encode) and keep the original text for display.

Expected result

Names such as "Zürich", "北京" or "Ærø" survive a round trip, sort and search as expected, and never produce a 500 because a header could not be encoded.

Limits and test basis

Normalisation and case folding are locale-independent approximations; Turkish dotted/dotless i and similar cases need locale-aware handling. The steps follow the cited references and a defect fixed in this wiki's own code.