v0.3.0

Added

• CONTRIBUTING.md: contributor workflow doc per the qte77 doc-structure canon — documentation hierarchy, commands, conventional commits, branches, changelog, and releasing (#75).

• CLAUDE.md: one-line @AGENTS.md import so the Claude Code loader and agents read the same rules (#75).

• .github/workflows/{bump-my-version,tag-release,publish-release}.yaml: estate-standard semi-automatic release pipeline (dispatch bump → PR → tag-on-merge → optional publish), mirrored from paperverse with SHA-pinned actions (#75).

• pyproject.toml: [tool.bumpversion] config + bump-my-version dev dependency, bumping pyproject.toml, src/pseudonymize_text/__init__.py, and the README version badge from a single source of truth (#75).

• Makefile: changelog_new / changelog_preview / changelog_release targets wrapping scriv (#75).

• --allow-broad-patterns CLI flag (detect/apply) — wires up the broad-pattern override that detectors.terms.load_terms already enforced but that was never exposed on the command line (docs/TERMS_CSV.md documented it).

• Opt-in phi detector group (--detectors phi) with checksum-validated NPI (Luhn over the 80840 prefix), DEA (registrant checksum), and VIN (ISO 3779 mod-11) detectors in detectors/phi.py; npi/dea/vin added to the default --types. Pure-Python, no new dependency. Part of #42 — clinical NER / MRN / date-coarsening remain deferred. New docs/PHI.md.

• detect and apply now print a one-line run summary to stderr on success — total spans, affected file count, and per-type counts (e.g. pseudonymize detect: 5 spans across 2 files — email:3 name:2). Counts only; never plaintext, context, or tokens. Part of D1.

• docs/decisions/README.md — an index of accepted ADRs (001–003) with status and date. Part of B2.

• Opt-in contextual MRN detection via --phi-context (with --detectors phi): a 6–10 digit run within 60 chars of an MRN cue (MRN, Medical Record Number, …) is tokenized as <MRN:…>. No checksum, so it is higher false-positive — review the report. Part of #42.

• Opt-in EU national-ID detectors via --detectors eu (tagged eu per ADR_003): German Steuer-ID (de_steuer), French NIR (fr_nir), UK NHS number (gb_nhs), Spanish DNI/NIE (es_dni), Italian Codice Fiscale (it_cf) — all checksum-validated via python-stdnum. Closes #86.

• examples/ — a runnable end-to-end demo (make demo) over a small mixed corpus spanning .txt/.md/.csv/.json/.log/.eml: public-domain excerpts (RFC 2822, a Lincoln letter) plus synthetic records that exercise every detector (structured + opt-in PHI + EU IDs, all checksum-valid) and mail handling (RFC 2047 header tokenization, DKIM-Signature strip, attachment drop). The demo uses an ephemeral key (live output gitignored under examples/_out/); examples/sample-output/ holds an illustrative committed snapshot. No key/mapping/plaintext committed. examples/README.md cross-references the larger doc-pipeline-engine corpus.

Changed

• README.md: restructured to the canonical estate order (Hero → Badges → What → How → Why → Refs → License); H1 renamed to pseudonymize-text, install fixed to uv add pseudonymize-text, Version badge linked to CHANGELOG.md, CodeQL badge added; static badges keep the teal 58f4c2 token by design (#75).

• AGENTS.md: ## Commands now points to CONTRIBUTING.md § Commands instead of duplicating the recipe list (#75).

• apply --plan re-detects .eml/.mbox parts (the plan keys spans by file and cannot be replayed across a message's MIME parts); detection is deterministic, so the result still matches the audited plan. Keep --terms on the apply command for mail corpora — documented in docs/USAGE.md.

• docs/landscape/de-identification.md: refreshed the comparison against verified upstream facts — added rows for philter-lite (maintained philter fork), Faker, and scrubadub; flagged philter-ucsf (last release 2020) and scrubadub (2023) as stale; corrected the scrubadub entry from "streaming" to library-API typed-placeholder redaction.

• tokenize.canonicalize and the report-record detector pattern now cover the PHI types; docs narrowed to reflect partial PHI support (README.md, docs/{COMPLIANCE,ARCHITECTURE,USAGE,USER_STORIES,HASHING,roadmap}.md, docs/landscape/de-identification.md).

• docs/USAGE.md: documented the previously-missing PSEUDONYMIZE_MAX_FILE_BYTES environment variable.

• Link checking now validates intra-document anchors (include_fragments in lychee.toml), so stale #fragment links in README.md/docs/ are caught in CI. Part of B1.

• markdownlint now also covers the root docs CONTRIBUTING.md, AGENTS.md, and CHANGELOG.md (in both make lint and the CI workflow). CLAUDE.md is excluded by omission — it is a one-line @AGENTS.md import that would otherwise trip MD041. Part of B3.

• Docs now tag detector coverage by jurisdiction (international / US / EU) and link the ADR_003 source-of-truth table from README, COMPLIANCE, GLOSSARY, USER_STORIES, and ARCHITECTURE (previously only USAGE linked it). Reconciled stale "MRN out of scope" wording across these docs (MRN is opt-in via --phi-context). Part of E4.

Fixed

• detect now scans .eml/.mbox mail (routed through formats/ like apply), so the audit plan is no longer silently empty for mail — closing a leak where the documented detect → apply --plan workflow shipped mail with PII intact.

• .eml/.mbox address headers (From/To/Cc/Bcc/Reply-To) are pseudonymised via getaddresses/formataddr, so display-name tokens are no longer mangled to <NAME>:hex> by the RFC 5322 address parser and the output re-parses cleanly.

• docs/USAGE.md: removed the non-existent --ner-confidence flag from the discovery-pass example (it was never implemented).

• detect now always writes the report header, even when zero spans are found, so a later apply --plan reads a valid empty plan (exit 0) instead of failing exit 4 on a missing file. apply emits the header for parity. Part of A1.

• apply --plan on a mail corpus with no --terms now warns on stderr: mail parts are re-detected with an empty term list, so literal entities (names/orgs) would silently pass through unredacted. Pass --terms for mail corpora. Part of A2.

• README install instructions: uv add pseudonymize-text failed because the package is not yet on PyPI. The quickstart now installs the CLI from git (uvx --from git+… / uv tool install git+…, PyPI publish planned) and is copy-pasteable (creates runs/in, a sample terms.csv, and the key). USAGE points to the README for install.

• Mail (.eml/.mbox) output is now deterministic. Pseudonymized headers are replaced in place instead of being deleted and re-added, which had re-ordered them in per-process (hash-randomized) frozenset order — so the same input + key now produces byte-identical output across runs, restoring the determinism guarantee for mail.