feat(knowledge-intake): classify a document + assess its impact before extraction

Phase A1. The real knowledge production is not writing — it is TARGETED UPDATING: when 20 documents
arrive, which 5 change our knowledge and which 15 are ignorable? Before the parser, Knowledge Intake
classifies a new document (no content extraction) and intersects its signals with an index of the
existing knowledge to emit a Knowledge Package (an impact analysis).

- compliance/knowledge_intake/: build_knowledge_index(patterns, playbooks, reference_scenarios,
  obligation_index) + assess_document_impact(descriptor, index) -> KnowledgePackage. Deterministic,
  NO content extraction, NO LLM. Surfaces affected capabilities / playbooks / transition patterns /
  reference scenarios / (injected) obligations, whether it is a new domain, and a triage level
  (HIGH / LOW / NONE / NEW_DOMAIN) with a recommendation.
- ADR-006: Knowledge Intake = classify + impact before extraction; full factory Intake -> Package ->
  Parser -> Draft -> Review -> Published; phase order A1 Intake / A2 Draft / A3 Review.
- reference suite: "Knowledge Intake" section triages 3 example documents (CRA SBOM-FAQ -> high,
  14C/2PB/3RTS/2Obl; environmental guidance -> new_domain; marketing blog -> ignorable). Section
  lives in _helpers.py to keep generate.py under the 500-LOC budget.
- Honest known refinement surfaced by intake: regulation-ID normalization (CRA vs Cyber Resilience Act).

10 intake tests (60 with the adjacent modules), mypy --strict clean (16 files), check-loc 0.
Product code with no app caller + ADR/reference = non-runtime -> no deploy (ADR-001). Freeze-safe.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend-compliance/reference_scenarios/_helpers.py

@@ -59,3 +59,48 @@ def unsupported_block(rmap) -> None:
 
 def interp_status(verdict_value: str) -> str:
     return "PARTIAL" if verdict_value in ("uncertain", "unsupported") else "PASS"
+
+
+def knowledge_intake_section(base_dir) -> None:
+    """Render the Knowledge Intake section (kept here so generate.py stays under the LOC budget)."""
+    import os
+    import yaml
+    from compliance.knowledge_intake import (
+        DocumentDescriptor, assess_document_impact, build_knowledge_index,
+    )
+
+    def _load(sub):
+        d = os.path.join(base_dir, "..", "knowledge", sub)
+        return [yaml.safe_load(open(os.path.join(d, f), encoding="utf-8"))
+                for f in sorted(os.listdir(d)) if f.endswith(".yaml")]
+
+    idx = build_knowledge_index(
+        _load("transition_patterns"), _load("implementation_playbooks"),
+        _load("reference_transition_scenarios"), obligation_index={"CRA": ["cra_obl_1", "cra_obl_2"]})
+    docs = [
+        DocumentDescriptor(document_id="ENISA CRA SBOM-FAQ", regulations=["CRA"], keywords=["sbom", "vulnerability"], document_type="faq"),
+        DocumentDescriptor(document_id="EU Umwelt-Leitfaden", regulations=["UmweltVO"], keywords=["wastewater"], document_type="guidance"),
+        DocumentDescriptor(document_id="Marketing-Blog", keywords=["newsletter"], document_type="blog"),
+    ]
+    w("## Knowledge Intake — Impact zuerst, Extraktion später")
+    w("")
+    w('_Vor dem Parser: ein neues Dokument NUR einordnen und seinen Impact auf den bestehenden Wissensbestand bestimmen. „Von N Dokumenten verändern wenige tatsächlich unser Wissen." Deterministisch, keine Extraktion, kein LLM._')
+    w("")
+    w("| Dokument | Impact | betrifft | Empfehlung |")
+    w("|---|---|---|---|")
+    for d in docs:
+        kp = assess_document_impact(d, idx)
+        touch = "neue Domäne" if kp.new_domain else "%dC·%dPB·%dRTS·%dObl" % (
+            len(kp.affected_capabilities), len(kp.affected_playbooks),
+            len(kp.affected_reference_scenarios), len(kp.affected_obligations))
+        w("| %s | **%s** | %s | %s |" % (d.document_id, kp.impact_level.value, touch, kp.recommendation.split(" —")[0]))
+    w("")
+    w("**Beispiel-Knowledge-Package** (`%s`): %s" % (docs[0].document_id, assess_document_impact(docs[0], idx).impact_summary))
+    w("")
+    w('_So entsteht bei jedem neuen Dokument eine Impact-Analyse statt „200 Seiten PDF" — Targeted Updating statt Schreiben._')
+    w("")
+    coverage_table([
+        ("Knowledge Intake (Klassifikation+Impact)", "PASS", "%d Regelwerke / %d Capabilities im Index" % (len(idx.regulations), len(idx.capability_regulations))),
+        ("Impact-Triage (HIGH/LOW/NONE/new_domain)", "PASS", "3 Beispiel-Dokumente korrekt eingeordnet"),
+        ("Regelwerk-ID-Normalisierung", "TODO", "CRA vs Cyber Resilience Act vereinheitlichen"),
+    ])