perf: preserve conversation prefixes across agent turns

tests/unit/agents/test_flash_runner_ledger.py

@@ -283,13 +283,13 @@ def test_injected_instruction_body_outlives_its_wrapper(mock_context):
     assert guidance.wrapper.startswith(USER_GUIDANCE_MARKER)
     assert "don't send messages" not in guidance.wrapper
 
-    # Scrub edge parity: once the turn has left the live position (a newer
-    # committed observation exists) the wrapper is gone and the body is
-    # still in the committed turn.
+    # Normal turns preserve guidance verbatim; an explicit compaction removes
+    # the old wrapper but preserves the user's standing instruction.
     ledger.stage_turn([tail, AIMessage(content="ok")])
     ledger.commit_staged(step_key="s1")
     ledger.stage_turn([runner._build_tail(ledger, 3, b"IMG", "[1] Send"), AIMessage(content="ok")])
     ledger.commit_staged(step_key="s2")
+    ledger.compact_active()
     rendered = ledger.render([runner._build_tail(ledger, 4, b"IMG", "[1] Send")])
     committed = next(
         m
@@ -679,8 +679,8 @@ async def test_run_calibrates_the_ledger_only_from_measured_usage(mock_context):
     assert record.call_count == 2
     first_call, second_call = record.call_args_list
     assert first_call.args[0] == 1200 and first_call.kwargs["messages"] is not None
-    # The report turn carries no usage metadata: estimated, recorded, not calibrated.
-    assert second_call.args[0] >= 1 and second_call.kwargs["messages"] is None
+    # Missing usage stays unknown; the next render computes its own estimate.
+    assert second_call.args[0] is None and second_call.kwargs["messages"] is None
 
 
 @pytest.mark.asyncio