0.5.0

.github/workflows/release.yml

@@ -99,6 +99,11 @@ jobs:
           # sdist separately for PyPI (model not included — source-only)
           uv build --sdist
 
+      - name: Verify PyPI metadata
+        run: |
+          pip install twine --quiet
+          twine check dist/*
+
       - name: Verify wheel install
         run: |
           python -m venv /tmp/verify-wheel

UNRELEASED.md

@@ -239,6 +239,7 @@
 - [TESTS]: Rewrite `test_summary.py` for `CombinedResult` schema
 - [ACTION]: Add `parse_result.py` — compute score, level, violations from `CombinedResult` JSON
 - [CLI]: `--strict` exits 1 on any finding, not just errors
+- [BUILD]: Remove `en-core-web-sm` direct URL dependency (rejected by PyPI). spaCy model auto-downloads on first run if missing.
 - [TESTS]: Fix remaining MCP and CLI integration tests for `CombinedResult` schema — remove `score`/`level` assertions, use `files`/`stats`
 
 ### Fixed

pyproject.toml

@@ -35,7 +35,6 @@ dependencies = [
     # hard requirement — the fallback verb lexicon drops charge accuracy
     # from 96% to ~78%. With the torch blocker active, spaCy loads in ~1s.
     "spacy>=3.8.11,<4",
-    "en-core-web-sm @ https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl",
 ]
 
 [project.optional-dependencies]
@@ -86,9 +85,6 @@ requires = ["hatchling"]
 build-backend = "hatchling.build"
 
 [tool.hatch.metadata]
-# Required because we pin en_core_web_sm via a direct URL (spaCy models
-# are distributed as GitHub release wheels, not via PyPI).
-allow-direct-references = true
 
 [tool.hatch.build.targets.wheel]
 packages = ["src/reporails_cli"]

src/reporails_cli/core/mapper/mapper.py

@@ -2091,10 +2091,24 @@ def nlp(self) -> Any | None:
                         # tok.text / tok.i / root.children. That needs tok2vec +
                         # tagger + parser only; ner / lemmatizer / attribute_ruler
                         # are dead weight on both load time and per-doc inference.
-                        self._nlp = spacy.load(
-                            "en_core_web_sm",
-                            disable=["ner", "lemmatizer", "attribute_ruler"],
-                        )
+                        try:
+                            self._nlp = spacy.load(
+                                "en_core_web_sm",
+                                disable=["ner", "lemmatizer", "attribute_ruler"],
+                            )
+                        except OSError:
+                            # Model not installed — download it once
+                            import subprocess
+                            import sys
+
+                            subprocess.run(
+                                [sys.executable, "-m", "spacy", "download", "en_core_web_sm"],
+                                capture_output=True,
+                            )
+                            self._nlp = spacy.load(
+                                "en_core_web_sm",
+                                disable=["ner", "lemmatizer", "attribute_ruler"],
+                            )
                     except (ImportError, OSError):
                         self._nlp = None
         return self._nlp