Python: Fix extract_range dropping multi-call assistant message and reordering parallel tool results (#14485) - #14778

Motivation and Context

Fixes #14485

extract_range() in python/semantic_kernel/contents/history_reducer/chat_history_reducer_utils.py is used by ChatHistorySummarizationReducer and ChatHistoryTruncationReducer (with preserve_pairs=True) to ensure that function calls and their corresponding results are kept or skipped together without leaving orphaned messages.

Previously, pair_map was constructed as a single flat dictionary:

pair_map = {}
for cidx, ridx in pairs:
    pair_map[cidx] = ridx
    pair_map[ridx] = cidx

When an assistant message contains multiple parallel tool calls (the standard shape for parallel/multi-function calling), all tool calls share the same call_index (cidx). Because pair_map[cidx] = ridx overwrote previous entries on each iteration, only the last result index was kept for cidx.

This caused two issues:

1. Orphaned tool results when filtering: If a filter_func was used to drop just one parallel tool result (e.g. moderation or specific error filtering), the call message evaluated its keep/skip condition solely against the single overwriting result in pair_map[cidx]. If that result happened to be the one filtered out, the entire assistant message was dropped even though sibling tool results were kept, producing orphaned AuthorRole.TOOL messages without a preceding assistant tool_calls message.
2. Scrambled chronological order: When iterating through the slice, eager appending of pair_map[idx] inserted the last result right after the call message, while earlier results were processed later in the loop, corrupting the chronological order of parallel tool results.

Description

• Replaced flat pair_map with call_to_results: dict[int, list[int]] (mapping each call index to all of its corresponding result indices) and result_to_call: dict[int, int].
• Updated extract_range() so that:
  • When evaluating an assistant call message, if only a subset of in-range parallel tool results is excluded by filter_func, the assistant call message is preserved alongside the surviving results.
  • If the call message itself is filtered out, or if all of its paired results are filtered out, both the call and its results are skipped.
  • Kept in-range results are sorted and deduplicated, guaranteeing that parallel tool results are appended in their original chronological order and preventing duplicate appends when a single TOOL message contains multiple FunctionResultContent items.
• Added comprehensive unit tests in tests/unit/contents/test_chat_history_reducer_utils.py:
  • test_extract_range_parallel_tool_calls_filter_partial_result: verifies that dropping one parallel result preserves the assistant call message and sibling result.
  • test_extract_range_parallel_tool_calls_preserves_chronological_order: verifies chronological ordering of parallel results.
  • test_extract_range_parallel_tool_calls_filter_call_message_drops_all_results: verifies that filtering the call message drops all paired results.
  • test_extract_range_parallel_tool_calls_filter_all_results_drops_call_message: verifies that filtering all results drops the call message.
  • test_extract_range_multiple_results_in_single_message: verifies correct handling of multiple results within a single message.

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the SK Contribution Guidelines and the pre-submission formatting script raises no violations
• All unit tests pass, and I have added new tests where possible
• I didn't break anyone 😄