Verified clean — compression END OF CONTEXT SUMMARY marker for all roles

Reviewed the marker propagation across all three insertion paths:

1. Standalone summary (summary_role == "user" or "assistant"): The condition changed from if not _merge_summary_into_tail and summary_role == "user" to if not _merge_summary_into_tail — correctly extending the marker to assistant-role standalone summaries. This addresses Context compression summary leaks into user-visible chat output #33256 where models regurgitate assistant-role summaries.

2. Merged-into-tail: The existing marker was inlined to use _SUMMARY_END_MARKER constant — no behavior change, just deduplication. Correct.

3. Rehydration stripping (_strip_summary_prefix): Added if text.endswith(_SUMMARY_END_MARKER): text = text[:-len(...)].rstrip() after prefix stripping. The break added to the prefix loop is correct — once a prefix matches, we skip the rest. Previously the code return-ed inside the loop, so the break + post-loop strip is equivalent but enables the marker stripping after the loop. The rstrip() preserves meaningful trailing whitespace from the summary body.

4. Test coverage: The new test_assistant_role_summary_carries_end_marker correctly constructs a head-last=user scenario that produces summary_role="assistant" and verifies both the role and the marker. The message sequence is carefully designed with protect_first_n=2, protect_last_n=2 and 8 messages to ensure the standalone insertion path.

No issues found.