fix: findings from the independent review

Scenario runner: stops when overridden base parameters leave the start out
of balance (new field startlikevekt: krev, kalibrer, ubalanse); stops on
non-finite results; says that models with a court queue drift slowly even
without a shock and prints the drift before the shock; queue in months of
inflow; never deletes files it did not write; stricter limits.
Check: anchored on the note's own checksums for core and reference files,
runs the scenario files too, exact comparison for the diagnosis, honest
wording when parts were skipped or an extension is broken.
Launcher: unknown field names in input are rejected with a suggestion;
unwritable folders give a message, not a traceback.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>

tests/conftest.py

@@ -70,10 +70,11 @@ def hjem(tmp_path: Path) -> Path:
 
 @pytest.fixture
 def kopi(tmp_path: Path) -> Path:
-    """En kopi av modell/, fasit/ og eksempler/ som testen kan endre. Returnerer kopiens rot."""
+    """En kopi av modell/, fasit/ og eksempler/ som testen kan endre. Returnerer kopiens rot.
+
+    Kopien er ikke merket som arbeidskopi (ingen LES-MEG.md), som en klone av repoet.
+    """
     rot = tmp_path / "kopi"
     for navn in ("modell", "fasit", "eksempler"):
         shutil.copytree(ROT / navn, rot / navn, ignore=shutil.ignore_patterns("__pycache__"))
-    for navn in ("scenario.py", "arbeidskopi.py"):
-        (rot / "modell" / navn).unlink(missing_ok=True)
     return rot