refactor: split flat core into capability packages, layer the conversations service, English defaults for every model-facing text

tests/test_claude_sdk_backend.py

@@ -32,8 +32,8 @@ from beaver_gateway.backends.claude_sdk import (
     UsageEvent,
     fingerprint,
 )
-from beaver_gateway.core.transcript import messages_from_entries
-from beaver_gateway.core.turn_capture import TurnCapture
+from beaver_gateway.backends.transcript import messages_from_entries
+from beaver_gateway.backends.capture import TurnCapture
 
 
 def _stream(index: int, text: str) -> list[StreamEvent]:
@@ -506,7 +506,7 @@ async def test_deltas_reach_the_caller_before_the_turn_ends(cwd: Path) -> None:
 
 
 async def test_policy_hook_denies_and_audits(cwd: Path) -> None:
-    from beaver_gateway.core.policy import Deny, ToolCall
+    from beaver_gateway.agents.policy import Deny, ToolCall
 
     def no_days(c: ToolCall):
         p = c.path()