Fix: map Anthropic compaction finish_reason to stop instead of length

When Anthropic's context compaction triggers with pause_after_compaction=True,
the API returns finish_reason "compaction". This was incorrectly mapped to
"length", which signals output truncation due to max_tokens and causes
downstream consumers to retry or attempt continuation. Compaction means the
context was summarized and the model paused naturally, so "stop" is the
semantically correct mapping.

Fixes #25165

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

litellm/litellm_core_utils/core_helpers.py

@@ -65,7 +65,7 @@ _FINISH_REASON_MAP: dict[str, OpenAIChatCompletionFinishReason] = {
     "max_tokens": "length",
     "tool_use": "tool_calls",
     "refusal": "content_filter",
-    "compaction": "length",
+    "compaction": "stop",
     # Cohere
     "COMPLETE": "stop",
     "ERROR_TOXIC": "content_filter",

tests/test_litellm/litellm_core_utils/test_core_helpers.py

@@ -66,7 +66,7 @@ class TestMapFinishReasonAnthropic:
             ("end_turn", "stop"),
             ("max_tokens", "length"),
             ("tool_use", "tool_calls"),
-            ("compaction", "length"),
+            ("compaction", "stop"),
             ("content_filtered", "content_filter"),
         ],
     )