[mostafa]

Fixes #594.

value_percentile and value_median correlations sorted every finite value in the group's window on each event fed into the group, and chain bookkeeping deep-cloned every fired correlation result only to read its source index and group key.

Changes

• Percentile and median select the needed ranks with select_nth_unstable_by(f64::total_cmp) instead of sorting. An interpolated percentile takes its upper neighbour as the minimum of the right-hand partition, and an even-count median takes its lower middle value as the maximum of the left-hand partition, so results are bit-identical to the sorted computation. check_condition and current_value now share window_percentile and window_median. Window state and snapshots are unchanged.
• chain_correlations carries (source_idx, group_key) in pending instead of full EvaluationResults, which removes the per-fire clone of every result (including retained events in full event mode) and the header clone for each chained parent.

Benchmark

New correlation_numeric_window group in benches/correlation.rs: one group held at a steady window size, 1,000 events per iteration, each checked against the full window.

Case	Before	After
value_percentile, 1,000 values	10.15 ms	3.00 ms
value_percentile, 5,000 values	42.75 ms	10.15 ms
value_median, 1,000 values	9.79 ms	2.68 ms
value_median, 5,000 values	42.55 ms	10.00 ms

Tests

• test_percentile_selection_matches_sorted_reference compares against a full-sort reference bit for bit, over window sizes 1 to 64 plus 99, 100, 1,000, and 5,001, at every integer rank and a few fractional ones. The samples include repeats and signed zeros.
• test_window_median_matches_sorted_reference does the same for odd and even medians through both check_condition and current_value.
• test_chaining_in_full_event_mode checks that a child carrying retained events still drives its parent, which fires with the child's group key and no events.
• The Sigma corpus regression is identical to main.