v0.5.0

Added

• Multi-file and directory input: the CLI now accepts one or more paths, and a
  directory is walked recursively for *.bib/*.md (hidden dirs like .git
  skipped). This makes the GitHub Action's default paths: . work as documented.
  When more than one file is scanned, findings are attributed with their filename
  (text path:Lnn, JSON file field). The pre-commit hook now passes staged
  filenames automatically (pass_filenames: true).

Fixed

• --fail-on now errors on an unknown tier name instead of silently dropping it
  (a typo like --fail-on retractions had turned the CI gate into a no-op).
• DOI cache round-trip no longer drops the preprint flags (is_preprint,
  has_preprint_relation); a preprint served from a warm cache is now downgraded
  identically to a cold lookup (restores determinism).
• Markdown/DOI-list entries (which carry no parsed title) no longer emit a
  misleading "possibly wrong DOI" verdict on an author mismatch; they fall back to
  the neutral "CrossRef first author is …" message.
• DOIs are URL-encoded before being spliced into CrossRef/OpenAlex/doi.org request
  URLs, so a DOI containing # (or other URL-significant characters) is no longer
  silently truncated into a different query.
• --json --dry-run now emits valid JSON instead of a plain-text line.

Packaging / CI

• Release workflow gains a test gate (build needs test), so a tag pushed on a
  red tree cannot publish. Action inputs are passed via env: (no shell
  interpolation), the Action's pip install is pinned, and the v1 major tag is
  force-moved to each release commit.
• .claude/ is git-ignored and excluded from the sdist (was leaking local
  worktrees into uv build artifacts). Added the Typing :: Typed classifier.
• test_semantic_live.py gains the live marker so the weekly live-tests
  workflow actually collects it (and that workflow now sets GHOSTCITE_LIVE=1).

v0.4.0

Added

• Venue-mismatch tier V (warn-only, opt-in): flags a cited journal/venue that disagrees with
  CrossRef's container-title, using abbreviation-tolerant matching (so "J. Mol. Biol." vs
  "Journal of Molecular Biology" does not fire). Informational by default; add --fail-on venue
  to gate on it. Suppressed when an author/title mismatch already fired.
• Preprint ↔ published awareness: a year difference that is explained by the DOI being a preprint
  (or having a preprint/published relation) is downgraded from a CI-failing year mismatch to a
  non-failing informational finding.
• Dead-DOI resolution probe: when a DOI is absent from CrossRef, a HEAD https://doi.org/<doi>
  distinguishes "dead/fabricated DOI" from "resolves but not in CrossRef" in the Tier U message.
  Best-effort, never fails CI on its own; --no-doi-probe skips it for fully offline runs.
• --badge <path>: writes a shields.io endpoint badge JSON (citation health) keyed to the active
  --fail-on threshold (green/clean vs red/N issues). Warn-only tiers do not turn it red.
• --cross-check now accepts a comma-separated list and adds openalex as a second source of
  truth (mirroring the PubMed pass: corroborate/conflict, raise findings CrossRef missed, and
  OR-combine retraction). --cross-check pubmed is unchanged; --cross-check pubmed,openalex
  runs both. Optional --openalex-mailto / OPENALEX_MAILTO for the polite pool.

v0.1.0

Initial release.

Added

• CrossRef author/year cross-check: compares each citation's claimed first-author
  surname and year against CrossRef's canonical record for the DOI.
• Input auto-detection for BibTeX, Markdown reference lists, and bare DOI lists
  (override with --format {auto,bibtex,markdown,doi}).
• Read the bibliography from stdin by passing - as the filename
  (e.g. cat refs.bib | ghostcite -).
• Severity tiers: A (author mismatch), B (year mismatch), C (cosmetic fold-only),
  R (retraction / expression-of-concern), U (unresolvable / not verifiable).
• Retraction and expression-of-concern detection via CrossRef updated-by /
  update-to / relation metadata.
• --cross-check pubmed: optional second source of truth via PubMed / NCBI
  E-utilities. Corroborates or conflicts CrossRef findings, can raise findings
  CrossRef missed, and supplies records for DOIs absent from CrossRef. Honors
  --ncbi-email / --ncbi-api-key (or NCBI_EMAIL / NCBI_API_KEY) for NCBI
  etiquette and a higher rate limit.
• Proactive rate pacing: self-throttles to CrossRef's advertised rate limit (read
  from response headers); --max-rps to cap further.
• --color {auto,always,never} colorized tier glyphs, honoring NO_COLOR.
• No-DOI entries resolved by best-effort bibliographic search, flagged
  low-confidence and never escalated above a warning on their own.
• --json machine-readable output, --dry-run parse-and-count (no network), and
  --fail-on CI gate to select which tiers force a non-zero exit.
• --version flag and python -m ghostcite module entry point.
• Composite GitHub Action (musharna/ghostcite@v1) for drop-in CI usage.
• Exit-code policy: 0 clean, 1 findings at/above the fail threshold, 2 tool
  error (network down, unparseable input).