fix(security): isolate template rendering with hard resource limits

docs/ADMIN_GUIDE.md

@@ -36,3 +36,21 @@ Apply the module Alembic migration before startup. Monitor rejected renders for
 contract drift, output limits, missing Files permission, and reused idempotency
 keys. HTML is designed for browser/OS printing; do not treat it as a signed PDF
 or proof of physical printer delivery.
+
+Pure rendering runs in a disposable Core worker after authorization, revision
+selection and idempotency lookup. Each worker has a 15-second wall limit,
+10 CPU seconds, 512 MiB address space and no file output. Typed input transport
+is limited to 32 MiB and result transport to 8 MiB, including serialization
+overhead; nesting is limited to 64 levels and 1,000,000 value nodes. The existing
+5 MiB final output and 5,000-item limits remain. Only data DTOs cross the
+boundary; principals, SQL sessions, credentials and artifact writes stay in the
+parent. Completed idempotent renders return before a worker is started.
+
+`GOVOPLAN_ISOLATED_PROCESS_CONCURRENCY` in Core controls active isolated work per
+API/worker process without queuing. Busy capacity produces a sanitized retryable
+render error. CPU, memory, transport or time failures stop the entire render
+before artifact persistence; reduce the workload before retrying. POSIX process
+resource controls are required. Missing controls, cancellation or worker failure
+fails closed with no in-process fallback. Budget aggregate memory across all
+slots and API/worker replicas; monitor limit and overload errors separately
+from template compatibility errors.

docs/USER_GUIDE.md

@@ -21,3 +21,13 @@ Open **Templates** to create or select a reusable definition.
 Render evidence shows the exact revision and abbreviated template, input, and
 output hashes. A consumer such as Campaign can submit many frozen recipients;
 the UI sample intentionally validates one representative item.
+
+Rendering is also limited to 15 seconds elapsed time, 10 CPU seconds, 512 MiB
+process memory, 32 MiB serialized input and 8 MiB serialized result transport;
+the final output remains limited to 5 MiB and 5,000 items. Deeply nested data
+or more than 1,000,000 transported value nodes is rejected. A limit error creates
+no partial output or render evidence: reduce the selected input or template
+complexity and retry. If processing capacity is busy, retry later. Unavailable
+worker controls or a worker failure require operator investigation. Existing
+successful output with the same idempotency key is reused without rendering
+again, provided its input and authorization still match.

src/govoplan_templates/backend/manifest.py

@@ -102,6 +102,43 @@ ROLE_TEMPLATES = (
 )
 
 DOCUMENTATION = (
+    DocumentationTopic(
+        id="templates.render-worker-limits",
+        title="Rendering resource and capacity limits",
+        summary="Render in a disposable process and reject incomplete output before persistence.",
+        body=(
+            "Template substitution and document composition run in a fresh process with a 15-second wall limit, 10 CPU seconds, "
+            "512 MiB address space, no file output, 32 MiB typed input and 8 MiB typed result transport. The existing final output "
+            "limit remains 5 MiB and the item limit remains 5,000. Transport permits only explicit data types, at most 64 nesting "
+            "levels and 1,000,000 value nodes. Exceeding a limit fails the entire render before Files or render evidence is written; "
+            "reduce the input or template complexity and retry. Core GOVOPLAN_ISOLATED_PROCESS_CONCURRENCY limits simultaneous work per "
+            "API/worker process without queuing; a busy error can be retried later. Missing POSIX resource controls, cancellation "
+            "and worker failures are explicit sanitized render errors, with no in-process fallback. Authorization, revision "
+            "selection, idempotency lookup, SQL sessions and artifact persistence remain in the parent. Existing completed "
+            "idempotent output is reused before starting a worker. Operators must budget aggregate memory across process slots."
+        ),
+        layer="static",
+        documentation_types=("user", "admin"),
+        audience=("user", "module_admin", "operator"),
+        order=30,
+        translations={"de": {
+            "title": "Ressourcen- und Kapazitätsgrenzen beim Rendern",
+            "summary": "In einem kurzlebigen Prozess rendern und unvollständige Ausgaben vor dem Speichern ablehnen.",
+            "body": (
+                "Platzhalterersetzung und Dokumentaufbau laufen in einem neuen Prozess mit höchstens 15 Sekunden Gesamtdauer, "
+                "10 CPU-Sekunden, 512 MiB Adressraum, keiner Dateiausgabe sowie 32 MiB typisierter Eingabe und 8 MiB Ergebnistransport. "
+                "Die endgültige Ausgabe bleibt auf 5 MiB und die Elementzahl auf 5.000 begrenzt. Der Transport erlaubt nur ausdrückliche "
+                "Datentypen, höchstens 64 Verschachtelungsebenen und 1.000.000 Wertknoten. Eine Grenzverletzung bricht den gesamten "
+                "Vorgang vor dem Speichern in Files oder als Render-Nachweis ab; reduzieren Sie Eingaben oder Vorlagenkomplexität "
+                "und versuchen Sie es erneut. Core GOVOPLAN_ISOLATED_PROCESS_CONCURRENCY begrenzt gleichzeitige Arbeiten pro API-/Worker-Prozess "
+                "ohne Warteschlange; bei belegter Kapazität kann später erneut versucht werden. Fehlende POSIX-Ressourcenbegrenzungen, "
+                "Abbruch und Worker-Fehler sind ausdrückliche bereinigte Render-Fehler ohne Rückfall auf den Elternprozess. "
+                "Berechtigungsprüfung, Revisionsauswahl, Idempotenzprüfung, SQL-Sitzungen und Artefaktspeicherung bleiben im Elternprozess. "
+                "Bereits abgeschlossene idempotente Ausgaben werden vor einem Worker-Start wiederverwendet. Betreiber müssen den "
+                "Gesamtspeicher aller Prozessplätze berücksichtigen."
+            ),
+        }},
+    ),
     DocumentationTopic(
         id="templates.workspace-layout",
         title="Templates workspace actions",

src/govoplan_templates/backend/rendering.py

@@ -25,6 +25,12 @@ from govoplan_core.core.templates import (
     TemplateRenderRequest,
     TemplateRenderResult,
 )
+from govoplan_core.security.bounded_process import (
+    ProcessBudgetError,
+    ProcessLimits,
+    run_bounded_operation,
+)
+from govoplan_core.security.worker_payload import decode_worker_payload, encode_worker_payload
 from govoplan_templates.backend.db.models import (
     TemplateDefinition,
     TemplateRender,
@@ -42,9 +48,36 @@ from govoplan_templates.backend.service import (
 RENDERER_VERSION = "templates-html-1"
 MAX_OUTPUT_BYTES = 5 * 1024 * 1024
 MAX_ITEMS = 5_000
+RENDER_PROCESS_LIMITS = ProcessLimits(
+    wall_seconds=15,
+    cpu_seconds=10,
+    memory_bytes=512 * 1024 * 1024,
+    input_bytes=32 * 1024 * 1024,
+    output_bytes=8 * 1024 * 1024,
+)
 _TOKEN_PATTERN = re.compile(r"{{\s*([A-Za-z_][A-Za-z0-9_.-]*)\s*}}")
 
 
+@dataclasses.dataclass(frozen=True, slots=True)
+class _RenderDefinition:
+    name: str
+
+
+@dataclasses.dataclass(frozen=True, slots=True)
+class _RenderRevision:
+    content_text: str | None
+    content_html: str | None
+    template_type: str
+    layout: dict[str, object]
+    output_profiles: list[dict[str, object]]
+
+
+@dataclasses.dataclass(frozen=True, slots=True)
+class _RenderRequest:
+    output_format: str
+    parameters: dict[str, object]
+
+
 class SqlTemplateRenderer:
     def __init__(self, registry: object | None = None) -> None:
         self.registry = registry
@@ -359,6 +392,68 @@ def _render_payload(
     *,
     request: TemplateRenderRequest,
     items: Sequence[Mapping[str, object]],
+) -> tuple[bytes, str, int]:
+    # Authoritative selection/idempotency precede this call. Transfer only the
+    # fields needed for pure rendering, never ORM objects, principals or Files.
+    try:
+        data = encode_worker_payload({
+            "definition": {"name": definition.name},
+            "revision": {
+                "content_text": revision.content_text,
+                "content_html": revision.content_html,
+                "template_type": revision.template_type,
+                "layout": dict(revision.layout),
+                "output_profiles": list(revision.output_profiles),
+            },
+            "request": {
+                "output_format": request.output_format,
+                "parameters": dict(request.parameters),
+            },
+            "items": tuple(dict(item) for item in items),
+        }, max_bytes=RENDER_PROCESS_LIMITS.input_bytes)
+        result = decode_worker_payload(
+            run_bounded_operation(_render_payload_worker, data, limits=RENDER_PROCESS_LIMITS),
+            max_bytes=RENDER_PROCESS_LIMITS.output_bytes,
+        )
+    except ProcessBudgetError as exc:
+        raise TemplateRenderError(f"Template rendering failed ({exc.code}): {exc}") from exc
+    except (TypeError, ValueError, RecursionError) as exc:
+        raise TemplateRenderError("Template input or output could not be safely transferred.") from exc
+    if not isinstance(result, dict):
+        raise TemplateRenderError("Template worker returned an invalid result.")
+    if "render_error" in result:
+        raise TemplateRenderError(str(result["render_error"]))
+    payload, content_type, page_count = result.get("payload"), result.get("content_type"), result.get("page_count")
+    if not isinstance(payload, bytes) or not isinstance(content_type, str) or type(page_count) is not int:
+        raise TemplateRenderError("Template worker returned an invalid result.")
+    if len(payload) > MAX_OUTPUT_BYTES:
+        _output_limit_exceeded()
+    return payload, content_type, page_count
+
+
+def _render_payload_worker(payload: bytes) -> bytes:
+    data = decode_worker_payload(payload, max_bytes=RENDER_PROCESS_LIMITS.input_bytes)
+    definition = _RenderDefinition(**data["definition"])
+    revision = _RenderRevision(**data["revision"])
+    request = _RenderRequest(**data["request"])
+    try:
+        payload, content_type, page_count = _render_payload_content(
+            definition, revision, request=request, items=data["items"]
+        )
+    except TemplateRenderError as exc:
+        return encode_worker_payload({"render_error": str(exc)})
+    return encode_worker_payload(
+        {"payload": payload, "content_type": content_type, "page_count": page_count},
+        max_bytes=RENDER_PROCESS_LIMITS.output_bytes,
+    )
+
+
+def _render_payload_content(
+    definition: _RenderDefinition,
+    revision: _RenderRevision,
+    *,
+    request: _RenderRequest,
+    items: Sequence[Mapping[str, object]],
 ) -> tuple[bytes, str, int]:
     if request.output_format == "text":
         body = revision.content_text or _html_to_text(revision.content_html or "")

tests/test_render_limits.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import unittest
+from dataclasses import replace
 from types import SimpleNamespace
 from unittest.mock import patch
 
@@ -47,6 +48,43 @@ class TemplateRenderLimitTests(unittest.TestCase):
             with self.assertRaises(TemplateRenderError):
                 rendering._render_payload(definition, revision, request=request, items=[{}])
 
+    def test_fresh_worker_preserves_exact_text_bytes_and_page_count(self):
+        definition = SimpleNamespace(name="Example")
+        revision = SimpleNamespace(content_html=None, content_text="Hi {{name}}: {{flag}}", template_type="serial_letter", layout={}, output_profiles=[])
+        request = SimpleNamespace(output_format="text", parameters={"flag": True})
+        with patch.object(rendering, "_render_payload_content", side_effect=AssertionError("parent renderer ran")):
+            result = rendering._render_payload(definition, revision, request=request, items=[{"name": "Äda"}, {"name": "Grace"}])
+        self.assertEqual(("Hi Äda: true\n\n---\n\nHi Grace: true".encode(), "text/plain; charset=utf-8", 2), result)
+
+    def test_real_worker_timeout_fails_without_parent_fallback(self):
+        definition = SimpleNamespace(name="Example")
+        revision = SimpleNamespace(content_html=None, content_text="Hello", template_type="serial_letter", layout={}, output_profiles=[])
+        request = SimpleNamespace(output_format="text", parameters={})
+        limits = replace(rendering.RENDER_PROCESS_LIMITS, wall_seconds=0.001)
+        with patch.object(rendering, "RENDER_PROCESS_LIMITS", limits), patch.object(
+            rendering, "_render_payload_content", side_effect=AssertionError("parent renderer ran")
+        ):
+            with self.assertRaisesRegex(TemplateRenderError, "timeout"):
+                rendering._render_payload(definition, revision, request=request, items=[{}])
+
+    def test_worker_html_bytes_match_existing_composition(self):
+        definition = SimpleNamespace(name="Letters <archive>")
+        revision = SimpleNamespace(content_html="<p>{{name}}: {{settings}}</p>", content_text=None, template_type="label_sheet", layout={"page_size": "A5", "columns": 2, "rows": 2}, output_profiles=[])
+        request = SimpleNamespace(output_format="html", parameters={"settings": {"b": True, "a": "ü"}})
+        items = [{"name": "<Ada>"}, {"name": "Grace & Co"}]
+        expected = rendering._render_payload_content(definition, revision, request=request, items=items)
+        actual = rendering._render_payload(definition, revision, request=request, items=items)
+        self.assertEqual(expected, actual)
+
+    def test_real_worker_output_transport_limit_is_enforced(self):
+        definition = SimpleNamespace(name="Example")
+        revision = SimpleNamespace(content_html=None, content_text="x" * 4096, template_type="serial_letter", layout={}, output_profiles=[])
+        request = SimpleNamespace(output_format="text", parameters={})
+        limits = replace(rendering.RENDER_PROCESS_LIMITS, output_bytes=1024)
+        with patch.object(rendering, "RENDER_PROCESS_LIMITS", limits):
+            with self.assertRaisesRegex(TemplateRenderError, "output_limit"):
+                rendering._render_payload(definition, revision, request=request, items=[{}])
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_templates.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import hashlib
 import unittest
+from dataclasses import replace
 from unittest.mock import patch
 
 from govoplan_core.auth import ApiPrincipal
@@ -19,6 +20,7 @@ from govoplan_core.core.templates import (
 )
 from govoplan_core.db.base import Base
 from govoplan_core.db.session import configure_database, reset_database
+from govoplan_templates.backend import rendering
 from govoplan_templates.backend.capabilities import (
     SqlTemplateCatalog,
     SqlTemplateContentLibrary,
@@ -231,6 +233,7 @@ class TemplateServiceTests(unittest.TestCase):
                 idempotency_key="campaign-1:postal-output-1",
             )
             first = render_template(session, principal(), registry=_Registry(), request=request)
+            with patch.object(rendering, "run_bounded_operation", side_effect=AssertionError("idempotent render started a worker")):
                 second = render_template(session, principal(), registry=_Registry(), request=request)
             session.commit()
 
@@ -243,6 +246,33 @@ class TemplateServiceTests(unittest.TestCase):
             self.assertIn(b"Ada", first.payload)
             self.assertIn(b"Grace", first.payload)
 
+    def test_real_worker_timeout_writes_neither_artifact_nor_render_evidence(self) -> None:
+        store = _ArtifactStore()
+        with self.database.session() as session:
+            item, _ = create_template(session, principal(), payload())
+            request = TemplateRenderRequest(
+                template_id=item.id,
+                usage="campaign.postal",
+                items=({"name": "Ada", "postal": {"address": "Street 1"}},),
+                persist_to_files=True,
+            )
+            limits = replace(rendering.RENDER_PROCESS_LIMITS, wall_seconds=0.001)
+            with patch.object(rendering, "RENDER_PROCESS_LIMITS", limits):
+                with self.assertRaisesRegex(TemplateRenderError, "timeout"):
+                    render_template(session, principal(), registry=_Registry(store), request=request)
+            self.assertIsNone(store.request)
+            self.assertEqual([], list_renders(session, principal()))
+
+    def test_incompatible_input_is_rejected_before_worker_admission(self) -> None:
+        with self.database.session() as session:
+            item, _ = create_template(session, principal(), payload())
+            request = TemplateRenderRequest(
+                template_id=item.id, usage="campaign.postal", items=({"name": "Missing address"},),
+            )
+            with patch.object(rendering, "run_bounded_operation", side_effect=AssertionError("incompatible render started a worker")):
+                with self.assertRaises(TemplateCompatibilityError):
+                    render_template(session, principal(), registry=_Registry(), request=request)
+
     def test_bounded_render_history_and_payload_are_owner_scoped(self) -> None:
         owner = principal(admin=False)
         other = principal(account_id="account-2", admin=False)