Complete tasks 3-7: harden verdict parsing, outputs retention, base branch, linux schedule parity, claim loop task

tests/test_vram_detect.py

@@ -84,7 +84,11 @@ def test_recommend_context_api_model() -> None:
     )
     assert recommended_kb > 0
     assert max_peak_kb > 0
-    assert recommended_kb <= 128_000 * 0.75  # after headroom
+    # Context sizing fix (task fix-context-sizing): headroom is applied
+    # EXACTLY ONCE. recommended_kb is the raw budget net of overhead
+    # (no headroom); max_peak_kb is post-headroom.
+    assert recommended_kb == 128_000 - 4000
+    assert max_peak_kb == (128_000 - 4000) * 75 // 100
 
 
 def test_recommend_context_manual_mode() -> None: