The catalog itself: every identifier as a PiiCatalog.Pattern.

This module is for maintainers and for readers who want to audit a regex. Use PiiCatalog.patterns/0 and the other functions in PiiCatalog to read the catalog.

Adding a pattern

1. An entry below. Its name is final once released, as callers store it. Use <identifier>_<country> in lowercase ASCII, the identifier in its local name and the country as its ISO 3166 code (personnummer_se, nino_gb); no country for an identifier used across borders (iban).
2. A validator in PiiCatalog.Validators where the identifier has a check digit or a structure the regex cannot express, with unit tests.
3. A vector file in test/vectors/<name>.exs with valid values, values with the right shape that fail the check, and look-alikes. The catalog test fails without one.
4. A benchmark with every pattern on, on prose and on number-heavy text, noted in the pull request. See "Scan cost" below.
5. A line in CHANGELOG.md naming the identifier, and the pii_catalog.readme Mix task run to regenerate the README table.

Boundaries

Patterns are compiled without the u flag so that they match bytewise and hold on text that is not valid UTF-8. \b then reads each byte as Latin-1, so the bytes of a non-ASCII character next to a value would count as word characters and the boundary would not be seen. The boundary fragments below ask only about ASCII word characters (A-Z, a-z, 0-9 and _): a value directly after or before any other byte, including the bytes of a non-ASCII letter, is found.

Where the pattern itself begins or ends with a word character, half of the boundary is already given and only the other half is asserted. The one lookaround keeps the scan several times cheaper than the two-sided form, which is needed only where the neighbouring character may be punctuation: the start of an email address.

Scan cost

PiiCatalog.Scanner scans every pattern that starts with (?<![A-Za-z0-9_]) and then a digit or a capital letter in one shared pass, so such a pattern adds little to the scan. A pattern that cannot start that way should start with a literal or a keyword with a single first letter (cpr, national), which the regex engine finds without trying every position; a keyword with several first letters, or a leading character class, costs a full pass of its own. A literal every match contains goes in requires, and the pattern is skipped on text without it. A keyword-led pattern whose keywords start with different letters lists one fast regex per letter in requires instead (svnr_at, nif_pt).