feat(tests): fixture sweep, anonymized real-invoice regression corpus, CI

- test_fixture_sweep: extract+validate every fixture PDF on every run;
  known extractor/validator gaps are catalogued explicitly and the sweep
  fails on any deviation from that catalogue
- tests/fixtures/real/: anonymized XML from real production invoices that
  revealed bugs (verified to contain no real-world data); first entry
  locks in the FC-tax-number/basis-quantity/spaced-VAT-ID regressions
- CI workflow: uv sync --frozen from lockfile + pytest + docker build

tests/test_fixture_sweep.py

@@ -0,0 +1,88 @@
+"""Regression sweep over all invoice fixtures.
+
+Every fixture PDF is extracted and validated end-to-end on every test run.
+Known limitations of extractor and validator are catalogued explicitly in
+KNOWN_GAPS and VALIDATION_EXCLUDED below. The sweep fails when:
+
+- any fixture stops extracting (exception),
+- a fixture that used to be clean produces new critical errors,
+- a fixture with known gaps produces more/different errors than catalogued,
+- a gap was fixed but its catalogue entry was not removed.
+
+So: fix a gap, then remove its entry — the sweep enforces honest bookkeeping.
+"""
+
+from pathlib import Path
+
+import pytest
+
+from src.extractor import extract_zugferd
+from src.models import ValidateRequest
+from src.validator import validate_invoice
+
+FIXTURES = Path(__file__).parent / "fixtures"
+ALL_CHECKS = ["pflichtfelder", "betraege", "ustid", "pdf_abgleich"]
+
+EXPECTED_NON_ZUGFERD = {"EmptyPDFA1.pdf"}
+
+VALIDATION_EXCLUDED = {
+    "ORDER-X_EX01_ORDER_FULL_DATA-COMFORT.pdf": (
+        "ORDER-X order document, not an invoice"
+    ),
+    "zugferd_invoice.pdf": (
+        "ZUGFeRD 1.0 uses a different XML namespace (not supported)"
+    ),
+}
+
+KNOWN_GAPS: dict[str, set[str]] = {
+    "EN16931_1_Teilrechnung.pdf": {
+        "line_items[3].line_total",
+        "totals.net",
+    },
+    "MustangBeispiel20221026.pdf": {
+        "totals.net",
+    },
+    "validAvoir_FR_type380_BASICWL.pdf": {
+        "line_items",
+        "totals.net",
+        "vat_id",
+    },
+    "zugferd_2p1_EXTENDED_PDFA-3A.pdf": {
+        "totals.net",
+    },
+}
+
+
+def _pdf_files() -> list[str]:
+    return sorted(path.name for path in FIXTURES.glob("*.pdf"))
+
+
+@pytest.mark.parametrize("filename", _pdf_files())
+def test_fixture_sweep(filename: str) -> None:
+    """Extract and validate one fixture; deviations from the catalogue fail."""
+    pdf_bytes = (FIXTURES / filename).read_bytes()
+    result = extract_zugferd(pdf_bytes)
+
+    if filename in EXPECTED_NON_ZUGFERD:
+        assert result.is_zugferd is False
+        return
+
+    assert result.is_zugferd is True, f"{filename}: expected ZUGFeRD XML"
+    assert result.xml_data is not None
+
+    if filename in VALIDATION_EXCLUDED:
+        return
+
+    request = ValidateRequest(
+        xml_data=result.xml_data.model_dump(),
+        pdf_text=result.pdf_text,
+        checks=ALL_CHECKS,
+    )
+    validation = validate_invoice(request)
+    critical_fields = {error.field for error in validation.errors}
+    expected_fields = KNOWN_GAPS.get(filename, set())
+
+    assert critical_fields == expected_fields, (
+        f"{filename}: critical fields changed — update KNOWN_GAPS "
+        f"(new state: {sorted(critical_fields)})"
+    )